Release prep: Phase 3 LM (checkpoints/phase3/p3_lm_full), versions/voices/settings, auto-tagger, Gradio app, CC-BY-SA-4.0 weights + Apache-2.0 code, model card

.gitattributes

@@ -40,3 +40,14 @@ prompts/clip3_waxal_yor_yor_1218.wav filter=lfs diff=lfs merge=lfs -text
 samples/prosody_v2/para_hau_eng_pcm.mp3 filter=lfs diff=lfs merge=lfs -text
 samples/prosody_v2/para_ibo_pcm_eng.mp3 filter=lfs diff=lfs merge=lfs -text
 samples/prosody_v2/para_yor_eng.mp3 filter=lfs diff=lfs merge=lfs -text
+assets/dynamic_accent_banner.jpg filter=lfs diff=lfs merge=lfs -text
+prompts/voice_waxal_ibo_5.wav filter=lfs diff=lfs merge=lfs -text
+prompts/voice_waxal_yor_5.wav filter=lfs diff=lfs merge=lfs -text
+samples/phase3/phase2/para_hau_eng_pcm.mp3 filter=lfs diff=lfs merge=lfs -text
+samples/phase3/phase2/para_ibo_pcm_eng.mp3 filter=lfs diff=lfs merge=lfs -text
+samples/phase3/phase2/para_yor_eng.mp3 filter=lfs diff=lfs merge=lfs -text
+samples/phase3/phase3/para_hau_eng_pcm.mp3 filter=lfs diff=lfs merge=lfs -text
+samples/phase3/phase3/para_ibo_pcm_eng.mp3 filter=lfs diff=lfs merge=lfs -text
+samples/phase3/phase3/para_yor_eng.mp3 filter=lfs diff=lfs merge=lfs -text
+samples/phase3/phase3_female/para_ibo_pcm_eng__waxal_ibo_5.mp3 filter=lfs diff=lfs merge=lfs -text
+samples/phase3/phase3_female/para_yor_eng__waxal_yor_5.mp3 filter=lfs diff=lfs merge=lfs -text

LICENSE

@@ -1,202 +1,428 @@
+Attribution-ShareAlike 4.0 International
+
+=======================================================================
+
+Creative Commons Corporation ("Creative Commons") is not a law firm and
+does not provide legal services or legal advice. Distribution of
+Creative Commons public licenses does not create a lawyer-client or
+other relationship. Creative Commons makes its licenses and related
+information available on an "as-is" basis. Creative Commons gives no
+warranties regarding its licenses, any material licensed under their
+terms and conditions, or any related information. Creative Commons
+disclaims all liability for damages resulting from their use to the
+fullest extent possible.
+
+Using Creative Commons Public Licenses
+
+Creative Commons public licenses provide a standard set of terms and
+conditions that creators and other rights holders may use to share
+original works of authorship and other material subject to copyright
+and certain other rights specified in the public license below. The
+following considerations are for informational purposes only, are not
+exhaustive, and do not form part of our licenses.
+
+     Considerations for licensors: Our public licenses are
+     intended for use by those authorized to give the public
+     permission to use material in ways otherwise restricted by
+     copyright and certain other rights. Our licenses are
+     irrevocable. Licensors should read and understand the terms
+     and conditions of the license they choose before applying it.
+     Licensors should also secure all rights necessary before
+     applying our licenses so that the public can reuse the
+     material as expected. Licensors should clearly mark any
+     material not subject to the license. This includes other CC-
+     licensed material, or material used under an exception or
+     limitation to copyright. More considerations for licensors:
+    wiki.creativecommons.org/Considerations_for_licensors
+
+     Considerations for the public: By using one of our public
+     licenses, a licensor grants the public permission to use the
+     licensed material under specified terms and conditions. If
+     the licensor's permission is not necessary for any reason--for
+     example, because of any applicable exception or limitation to
+     copyright--then that use is not regulated by the license. Our
+     licenses grant only permissions under copyright and certain
+     other rights that a licensor has authority to grant. Use of
+     the licensed material may still be restricted for other
+     reasons, including because others have copyright or other
+     rights in the material. A licensor may make special requests,
+     such as asking that all changes be marked or described.
+     Although not required by our licenses, you are encouraged to
+     respect those requests where reasonable. More considerations
+     for the public:
+    wiki.creativecommons.org/Considerations_for_licensees
+
+=======================================================================
+
+Creative Commons Attribution-ShareAlike 4.0 International Public
+License
+
+By exercising the Licensed Rights (defined below), You accept and agree
+to be bound by the terms and conditions of this Creative Commons
+Attribution-ShareAlike 4.0 International Public License ("Public
+License"). To the extent this Public License may be interpreted as a
+contract, You are granted the Licensed Rights in consideration of Your
+acceptance of these terms and conditions, and the Licensor grants You
+such rights in consideration of benefits the Licensor receives from
+making the Licensed Material available under these terms and
+conditions.
+
+
+Section 1 -- Definitions.
+
+  a. Adapted Material means material subject to Copyright and Similar
+     Rights that is derived from or based upon the Licensed Material
+     and in which the Licensed Material is translated, altered,
+     arranged, transformed, or otherwise modified in a manner requiring
+     permission under the Copyright and Similar Rights held by the
+     Licensor. For purposes of this Public License, where the Licensed
+     Material is a musical work, performance, or sound recording,
+     Adapted Material is always produced where the Licensed Material is
+     synched in timed relation with a moving image.
+
+  b. Adapter's License means the license You apply to Your Copyright
+     and Similar Rights in Your contributions to Adapted Material in
+     accordance with the terms and conditions of this Public License.
+
+  c. BY-SA Compatible License means a license listed at
+     creativecommons.org/compatiblelicenses, approved by Creative
+     Commons as essentially the equivalent of this Public License.
+
+  d. Copyright and Similar Rights means copyright and/or similar rights
+     closely related to copyright including, without limitation,
+     performance, broadcast, sound recording, and Sui Generis Database
+     Rights, without regard to how the rights are labeled or
+     categorized. For purposes of this Public License, the rights
+     specified in Section 2(b)(1)-(2) are not Copyright and Similar
+     Rights.
+
+  e. Effective Technological Measures means those measures that, in the
+     absence of proper authority, may not be circumvented under laws
+     fulfilling obligations under Article 11 of the WIPO Copyright
+     Treaty adopted on December 20, 1996, and/or similar international
+     agreements.
+
+  f. Exceptions and Limitations means fair use, fair dealing, and/or
+     any other exception or limitation to Copyright and Similar Rights
+     that applies to Your use of the Licensed Material.
+
+  g. License Elements means the license attributes listed in the name
+     of a Creative Commons Public License. The License Elements of this
+     Public License are Attribution and ShareAlike.
+
+  h. Licensed Material means the artistic or literary work, database,
+     or other material to which the Licensor applied this Public
+     License.
+
+  i. Licensed Rights means the rights granted to You subject to the
+     terms and conditions of this Public License, which are limited to
+     all Copyright and Similar Rights that apply to Your use of the
+     Licensed Material and that the Licensor has authority to license.
+
+  j. Licensor means the individual(s) or entity(ies) granting rights
+     under this Public License.
+
+  k. Share means to provide material to the public by any means or
+     process that requires permission under the Licensed Rights, such
+     as reproduction, public display, public performance, distribution,
+     dissemination, communication, or importation, and to make material
+     available to the public including in ways that members of the
+     public may access the material from a place and at a time
+     individually chosen by them.
+
+  l. Sui Generis Database Rights means rights other than copyright
+     resulting from Directive 96/9/EC of the European Parliament and of
+     the Council of 11 March 1996 on the legal protection of databases,
+     as amended and/or succeeded, as well as other essentially
+     equivalent rights anywhere in the world.
+
+  m. You means the individual or entity exercising the Licensed Rights
+     under this Public License. Your has a corresponding meaning.
+
+
+Section 2 -- Scope.
+
+  a. License grant.
+
+       1. Subject to the terms and conditions of this Public License,
+          the Licensor hereby grants You a worldwide, royalty-free,
+          non-sublicensable, non-exclusive, irrevocable license to
+          exercise the Licensed Rights in the Licensed Material to:
+
+            a. reproduce and Share the Licensed Material, in whole or
+               in part; and
+
+            b. produce, reproduce, and Share Adapted Material.
+
+       2. Exceptions and Limitations. For the avoidance of doubt, where
+          Exceptions and Limitations apply to Your use, this Public
+          License does not apply, and You do not need to comply with
+          its terms and conditions.
+
+       3. Term. The term of this Public License is specified in Section
+          6(a).
+
+       4. Media and formats; technical modifications allowed. The
+          Licensor authorizes You to exercise the Licensed Rights in
+          all media and formats whether now known or hereafter created,
+          and to make technical modifications necessary to do so. The
+          Licensor waives and/or agrees not to assert any right or
+          authority to forbid You from making technical modifications
+          necessary to exercise the Licensed Rights, including
+          technical modifications necessary to circumvent Effective
+          Technological Measures. For purposes of this Public License,
+          simply making modifications authorized by this Section 2(a)
+          (4) never produces Adapted Material.
+
+       5. Downstream recipients.
+
+            a. Offer from the Licensor -- Licensed Material. Every
+               recipient of the Licensed Material automatically
+               receives an offer from the Licensor to exercise the
+               Licensed Rights under the terms and conditions of this
+               Public License.
+
+            b. Additional offer from the Licensor -- Adapted Material.
+               Every recipient of Adapted Material from You
+               automatically receives an offer from the Licensor to
+               exercise the Licensed Rights in the Adapted Material
+               under the conditions of the Adapter's License You apply.
+
+            c. No downstream restrictions. You may not offer or impose
+               any additional or different terms or conditions on, or
+               apply any Effective Technological Measures to, the
+               Licensed Material if doing so restricts exercise of the
+               Licensed Rights by any recipient of the Licensed
+               Material.
+
+       6. No endorsement. Nothing in this Public License constitutes or
+          may be construed as permission to assert or imply that You
+          are, or that Your use of the Licensed Material is, connected
+          with, or sponsored, endorsed, or granted official status by,
+          the Licensor or others designated to receive attribution as
+          provided in Section 3(a)(1)(A)(i).
+
+  b. Other rights.
+
+       1. Moral rights, such as the right of integrity, are not
+          licensed under this Public License, nor are publicity,
+          privacy, and/or other similar personality rights; however, to
+          the extent possible, the Licensor waives and/or agrees not to
+          assert any such rights held by the Licensor to the limited
+          extent necessary to allow You to exercise the Licensed
+          Rights, but not otherwise.
+
+       2. Patent and trademark rights are not licensed under this
+          Public License.
+
+       3. To the extent possible, the Licensor waives any right to
+          collect royalties from You for the exercise of the Licensed
+          Rights, whether directly or through a collecting society
+          under any voluntary or waivable statutory or compulsory
+          licensing scheme. In all other cases the Licensor expressly
+          reserves any right to collect such royalties.
+
+
+Section 3 -- License Conditions.
+
+Your exercise of the Licensed Rights is expressly made subject to the
+following conditions.
+
+  a. Attribution.
+
+       1. If You Share the Licensed Material (including in modified
+          form), You must:
+
+            a. retain the following if it is supplied by the Licensor
+               with the Licensed Material:
+
+                 i. identification of the creator(s) of the Licensed
+                    Material and any others designated to receive
+                    attribution, in any reasonable manner requested by
+                    the Licensor (including by pseudonym if
+                    designated);
+
+                ii. a copyright notice;
+
+               iii. a notice that refers to this Public License;
+
+                iv. a notice that refers to the disclaimer of
+                    warranties;
+
+                 v. a URI or hyperlink to the Licensed Material to the
+                    extent reasonably practicable;
+
+            b. indicate if You modified the Licensed Material and
+               retain an indication of any previous modifications; and
+
+            c. indicate the Licensed Material is licensed under this
+               Public License, and include the text of, or the URI or
+               hyperlink to, this Public License.
+
+       2. You may satisfy the conditions in Section 3(a)(1) in any
+          reasonable manner based on the medium, means, and context in
+          which You Share the Licensed Material. For example, it may be
+          reasonable to satisfy the conditions by providing a URI or
+          hyperlink to a resource that includes the required
+          information.
+
+       3. If requested by the Licensor, You must remove any of the
+          information required by Section 3(a)(1)(A) to the extent
+          reasonably practicable.
+
+  b. ShareAlike.
+
+     In addition to the conditions in Section 3(a), if You Share
+     Adapted Material You produce, the following conditions also apply.
+
+       1. The Adapter's License You apply must be a Creative Commons
+          license with the same License Elements, this version or
+          later, or a BY-SA Compatible License.
+
+       2. You must include the text of, or the URI or hyperlink to, the
+          Adapter's License You apply. You may satisfy this condition
+          in any reasonable manner based on the medium, means, and
+          context in which You Share Adapted Material.
+
+       3. You may not offer or impose any additional or different terms
+          or conditions on, or apply any Effective Technological
+          Measures to, Adapted Material that restrict exercise of the
+          rights granted under the Adapter's License You apply.
+
+
+Section 4 -- Sui Generis Database Rights.
+
+Where the Licensed Rights include Sui Generis Database Rights that
+apply to Your use of the Licensed Material:
+
+  a. for the avoidance of doubt, Section 2(a)(1) grants You the right
+     to extract, reuse, reproduce, and Share all or a substantial
+     portion of the contents of the database;
+
+  b. if You include all or a substantial portion of the database
+     contents in a database in which You have Sui Generis Database
+     Rights, then the database in which You have Sui Generis Database
+     Rights (but not its individual contents) is Adapted Material,
+     including for purposes of Section 3(b); and
+
+  c. You must comply with the conditions in Section 3(a) if You Share
+     all or a substantial portion of the contents of the database.
+
+For the avoidance of doubt, this Section 4 supplements and does not
+replace Your obligations under this Public License where the Licensed
+Rights include other Copyright and Similar Rights.
+
+
+Section 5 -- Disclaimer of Warranties and Limitation of Liability.
+
+  a. UNLESS OTHERWISE SEPARATELY UNDERTAKEN BY THE LICENSOR, TO THE
+     EXTENT POSSIBLE, THE LICENSOR OFFERS THE LICENSED MATERIAL AS-IS
+     AND AS-AVAILABLE, AND MAKES NO REPRESENTATIONS OR WARRANTIES OF
+     ANY KIND CONCERNING THE LICENSED MATERIAL, WHETHER EXPRESS,
+     IMPLIED, STATUTORY, OR OTHER. THIS INCLUDES, WITHOUT LIMITATION,
+     WARRANTIES OF TITLE, MERCHANTABILITY, FITNESS FOR A PARTICULAR
+     PURPOSE, NON-INFRINGEMENT, ABSENCE OF LATENT OR OTHER DEFECTS,
+     ACCURACY, OR THE PRESENCE OR ABSENCE OF ERRORS, WHETHER OR NOT
+     KNOWN OR DISCOVERABLE. WHERE DISCLAIMERS OF WARRANTIES ARE NOT
+     ALLOWED IN FULL OR IN PART, THIS DISCLAIMER MAY NOT APPLY TO YOU.
+
+  b. TO THE EXTENT POSSIBLE, IN NO EVENT WILL THE LICENSOR BE LIABLE
+     TO YOU ON ANY LEGAL THEORY (INCLUDING, WITHOUT LIMITATION,
+     NEGLIGENCE) OR OTHERWISE FOR ANY DIRECT, SPECIAL, INDIRECT,
+     INCIDENTAL, CONSEQUENTIAL, PUNITIVE, EXEMPLARY, OR OTHER LOSSES,
+     COSTS, EXPENSES, OR DAMAGES ARISING OUT OF THIS PUBLIC LICENSE OR
+     USE OF THE LICENSED MATERIAL, EVEN IF THE LICENSOR HAS BEEN
+     ADVISED OF THE POSSIBILITY OF SUCH LOSSES, COSTS, EXPENSES, OR
+     DAMAGES. WHERE A LIMITATION OF LIABILITY IS NOT ALLOWED IN FULL OR
+     IN PART, THIS LIMITATION MAY NOT APPLY TO YOU.
+
+  c. The disclaimer of warranties and limitation of liability provided
+     above shall be interpreted in a manner that, to the extent
+     possible, most closely approximates an absolute disclaimer and
+     waiver of all liability.
+
+
+Section 6 -- Term and Termination.
+
+  a. This Public License applies for the term of the Copyright and
+     Similar Rights licensed here. However, if You fail to comply with
+     this Public License, then Your rights under this Public License
+     terminate automatically.
+
+  b. Where Your right to use the Licensed Material has terminated under
+     Section 6(a), it reinstates:
+
+       1. automatically as of the date the violation is cured, provided
+          it is cured within 30 days of Your discovery of the
+          violation; or
+
+       2. upon express reinstatement by the Licensor.
+
+     For the avoidance of doubt, this Section 6(b) does not affect any
+     right the Licensor may have to seek remedies for Your violations
+     of this Public License.
+
+  c. For the avoidance of doubt, the Licensor may also offer the
+     Licensed Material under separate terms or conditions or stop
+     distributing the Licensed Material at any time; however, doing so
+     will not terminate this Public License.
+
+  d. Sections 1, 5, 6, 7, and 8 survive termination of this Public
+     License.
+
+
+Section 7 -- Other Terms and Conditions.
+
+  a. The Licensor shall not be bound by any additional or different
+     terms or conditions communicated by You unless expressly agreed.
+
+  b. Any arrangements, understandings, or agreements regarding the
+     Licensed Material not stated herein are separate from and
+     independent of the terms and conditions of this Public License.
+
+
+Section 8 -- Interpretation.
+
+  a. For the avoidance of doubt, this Public License does not, and
+     shall not be interpreted to, reduce, limit, restrict, or impose
+     conditions on any use of the Licensed Material that could lawfully
+     be made without permission under this Public License.
+
+  b. To the extent possible, if any provision of this Public License is
+     deemed unenforceable, it shall be automatically reformed to the
+     minimum extent necessary to make it enforceable. If the provision
+     cannot be reformed, it shall be severed from this Public License
+     without affecting the enforceability of the remaining terms and
+     conditions.
+
+  c. No term or condition of this Public License will be waived and no
+     failure to comply consented to unless expressly agreed to by the
+     Licensor.
+
+  d. Nothing in this Public License constitutes or may be interpreted
+     as a limitation upon, or waiver of, any privileges and immunities
+     that apply to the Licensor or You, including from the legal
+     processes of any jurisdiction or authority.
+
+
+=======================================================================
+
+Creative Commons is not a party to its public
+licenses. Notwithstanding, Creative Commons may elect to apply one of
+its public licenses to material it publishes and in those instances
+will be considered the “Licensor.” The text of the Creative Commons
+public licenses is dedicated to the public domain under the CC0 Public
+Domain Dedication. Except for the limited purpose of indicating that
+material is shared under a Creative Commons public license or as
+otherwise permitted by the Creative Commons policies published at
+creativecommons.org/policies, Creative Commons does not authorize the
+use of the trademark "Creative Commons" or any other trademark or logo
+of Creative Commons without its prior written consent including,
+without limitation, in connection with any unauthorized modifications
+to any of its public licenses or any other arrangements,
+understandings, or agreements concerning use of licensed material. For
+the avoidance of doubt, this paragraph does not form part of the
+public licenses.
+
+Creative Commons may be contacted at creativecommons.org.
 
-                                 Apache License
-                           Version 2.0, January 2004
-                        http://www.apache.org/licenses/
-
-   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
-
-   1. Definitions.
-
-      "License" shall mean the terms and conditions for use, reproduction,
-      and distribution as defined by Sections 1 through 9 of this document.
-
-      "Licensor" shall mean the copyright owner or entity authorized by
-      the copyright owner that is granting the License.
-
-      "Legal Entity" shall mean the union of the acting entity and all
-      other entities that control, are controlled by, or are under common
-      control with that entity. For the purposes of this definition,
-      "control" means (i) the power, direct or indirect, to cause the
-      direction or management of such entity, whether by contract or
-      otherwise, or (ii) ownership of fifty percent (50%) or more of the
-      outstanding shares, or (iii) beneficial ownership of such entity.
-
-      "You" (or "Your") shall mean an individual or Legal Entity
-      exercising permissions granted by this License.
-
-      "Source" form shall mean the preferred form for making modifications,
-      including but not limited to software source code, documentation
-      source, and configuration files.
-
-      "Object" form shall mean any form resulting from mechanical
-      transformation or translation of a Source form, including but
-      not limited to compiled object code, generated documentation,
-      and conversions to other media types.
-
-      "Work" shall mean the work of authorship, whether in Source or
-      Object form, made available under the License, as indicated by a
-      copyright notice that is included in or attached to the work
-      (an example is provided in the Appendix below).
-
-      "Derivative Works" shall mean any work, whether in Source or Object
-      form, that is based on (or derived from) the Work and for which the
-      editorial revisions, annotations, elaborations, or other modifications
-      represent, as a whole, an original work of authorship. For the purposes
-      of this License, Derivative Works shall not include works that remain
-      separable from, or merely link (or bind by name) to the interfaces of,
-      the Work and Derivative Works thereof.
-
-      "Contribution" shall mean any work of authorship, including
-      the original version of the Work and any modifications or additions
-      to that Work or Derivative Works thereof, that is intentionally
-      submitted to Licensor for inclusion in the Work by the copyright owner
-      or by an individual or Legal Entity authorized to submit on behalf of
-      the copyright owner. For the purposes of this definition, "submitted"
-      means any form of electronic, verbal, or written communication sent
-      to the Licensor or its representatives, including but not limited to
-      communication on electronic mailing lists, source code control systems,
-      and issue tracking systems that are managed by, or on behalf of, the
-      Licensor for the purpose of discussing and improving the Work, but
-      excluding communication that is conspicuously marked or otherwise
-      designated in writing by the copyright owner as "Not a Contribution."
-
-      "Contributor" shall mean Licensor and any individual or Legal Entity
-      on behalf of whom a Contribution has been received by Licensor and
-      subsequently incorporated within the Work.
-
-   2. Grant of Copyright License. Subject to the terms and conditions of
-      this License, each Contributor hereby grants to You a perpetual,
-      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
-      copyright license to reproduce, prepare Derivative Works of,
-      publicly display, publicly perform, sublicense, and distribute the
-      Work and such Derivative Works in Source or Object form.
-
-   3. Grant of Patent License. Subject to the terms and conditions of
-      this License, each Contributor hereby grants to You a perpetual,
-      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
-      (except as stated in this section) patent license to make, have made,
-      use, offer to sell, sell, import, and otherwise transfer the Work,
-      where such license applies only to those patent claims licensable
-      by such Contributor that are necessarily infringed by their
-      Contribution(s) alone or by combination of their Contribution(s)
-      with the Work to which such Contribution(s) was submitted. If You
-      institute patent litigation against any entity (including a
-      cross-claim or counterclaim in a lawsuit) alleging that the Work
-      or a Contribution incorporated within the Work constitutes direct
-      or contributory patent infringement, then any patent licenses
-      granted to You under this License for that Work shall terminate
-      as of the date such litigation is filed.
-
-   4. Redistribution. You may reproduce and distribute copies of the
-      Work or Derivative Works thereof in any medium, with or without
-      modifications, and in Source or Object form, provided that You
-      meet the following conditions:
-
-      (a) You must give any other recipients of the Work or
-          Derivative Works a copy of this License; and
-
-      (b) You must cause any modified files to carry prominent notices
-          stating that You changed the files; and
-
-      (c) You must retain, in the Source form of any Derivative Works
-          that You distribute, all copyright, patent, trademark, and
-          attribution notices from the Source form of the Work,
-          excluding those notices that do not pertain to any part of
-          the Derivative Works; and
-
-      (d) If the Work includes a "NOTICE" text file as part of its
-          distribution, then any Derivative Works that You distribute must
-          include a readable copy of the attribution notices contained
-          within such NOTICE file, excluding those notices that do not
-          pertain to any part of the Derivative Works, in at least one
-          of the following places: within a NOTICE text file distributed
-          as part of the Derivative Works; within the Source form or
-          documentation, if provided along with the Derivative Works; or,
-          within a display generated by the Derivative Works, if and
-          wherever such third-party notices normally appear. The contents
-          of the NOTICE file are for informational purposes only and
-          do not modify the License. You may add Your own attribution
-          notices within Derivative Works that You distribute, alongside
-          or as an addendum to the NOTICE text from the Work, provided
-          that such additional attribution notices cannot be construed
-          as modifying the License.
-
-      You may add Your own copyright statement to Your modifications and
-      may provide additional or different license terms and conditions
-      for use, reproduction, or distribution of Your modifications, or
-      for any such Derivative Works as a whole, provided Your use,
-      reproduction, and distribution of the Work otherwise complies with
-      the conditions stated in this License.
-
-   5. Submission of Contributions. Unless You explicitly state otherwise,
-      any Contribution intentionally submitted for inclusion in the Work
-      by You to the Licensor shall be under the terms and conditions of
-      this License, without any additional terms or conditions.
-      Notwithstanding the above, nothing herein shall supersede or modify
-      the terms of any separate license agreement you may have executed
-      with Licensor regarding such Contributions.
-
-   6. Trademarks. This License does not grant permission to use the trade
-      names, trademarks, service marks, or product names of the Licensor,
-      except as required for reasonable and customary use in describing the
-      origin of the Work and reproducing the content of the NOTICE file.
-
-   7. Disclaimer of Warranty. Unless required by applicable law or
-      agreed to in writing, Licensor provides the Work (and each
-      Contributor provides its Contributions) on an "AS IS" BASIS,
-      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
-      implied, including, without limitation, any warranties or conditions
-      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
-      PARTICULAR PURPOSE. You are solely responsible for determining the
-      appropriateness of using or redistributing the Work and assume any
-      risks associated with Your exercise of permissions under this License.
-
-   8. Limitation of Liability. In no event and under no legal theory,
-      whether in tort (including negligence), contract, or otherwise,
-      unless required by applicable law (such as deliberate and grossly
-      negligent acts) or agreed to in writing, shall any Contributor be
-      liable to You for damages, including any direct, indirect, special,
-      incidental, or consequential damages of any character arising as a
-      result of this License or out of the use or inability to use the
-      Work (including but not limited to damages for loss of goodwill,
-      work stoppage, computer failure or malfunction, or any and all
-      other commercial damages or losses), even if such Contributor
-      has been advised of the possibility of such damages.
-
-   9. Accepting Warranty or Additional Liability. While redistributing
-      the Work or Derivative Works thereof, You may choose to offer,
-      and charge a fee for, acceptance of support, warranty, indemnity,
-      or other liability obligations and/or rights consistent with this
-      License. However, in accepting such obligations, You may act only
-      on Your own behalf and on Your sole responsibility, not on behalf
-      of any other Contributor, and only if You agree to indemnify,
-      defend, and hold each Contributor harmless for any liability
-      incurred by, or claims asserted against, such Contributor by reason
-      of your accepting any such warranty or additional liability.
-
-   END OF TERMS AND CONDITIONS
-
-   APPENDIX: How to apply the Apache License to your work.
-
-      To apply the Apache License to your work, attach the following
-      boilerplate notice, with the fields enclosed by brackets "[]"
-      replaced with your own identifying information. (Don't include
-      the brackets!)  The text should be enclosed in the appropriate
-      comment syntax for the file format. We also recommend that a
-      file or class name and description of purpose be included on the
-      same "printed page" as the copyright notice for easier
-      identification within third-party archives.
-
-   Copyright [yyyy] [name of copyright owner]
-
-   Licensed under the Apache License, Version 2.0 (the "License");
-   you may not use this file except in compliance with the License.
-   You may obtain a copy of the License at
-
-       http://www.apache.org/licenses/LICENSE-2.0
-
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-   See the License for the specific language governing permissions and
-   limitations under the License.

LICENSE-CODE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

README.md

@@ -1,5 +1,5 @@
 ---
-license: apache-2.0
+license: cc-by-sa-4.0
 language:
   - en
   - pcm
@@ -8,177 +8,277 @@ language:
   - ha
 base_model: FunAudioLLM/CosyVoice2-0.5B
 pipeline_tag: text-to-speech
+library_name: pytorch
 tags:
   - text-to-speech
   - code-switching
   - nigerian-languages
+  - igbo
+  - yoruba
+  - hausa
+  - nigerian-pidgin
   - cosyvoice2
-  - research-preview
+datasets:
+  - google/WaxalNLP
+  - McGill-NLP/NaijaS2ST
+  - AfriSpeech/open-bible-speech-african
+  - nkowaokwu/ibo-dict
+  - google/fleurs
 ---
 
-# Dynamic Accent (Phase 2, research preview)
+<p align="center"><img src="assets/dynamic_accent_banner.jpg" width="600"/></p>
 
-Nigerian multilingual, **code-switching** text-to-speech in one consistent voice: English, Nigerian Pidgin, Yoruba, Igbo and Hausa, mixed freely inside one sentence.
+# Dynamic Accent: Code-switching TTS
 
-It is built on [FunAudioLLM/CosyVoice2-0.5B](https://huggingface.co/FunAudioLLM/CosyVoice2-0.5B) and adds:
-- per-token **language embeddings**, a **speaker token** and an adversarial (GRL) speaker classifier on the LM, fine-tuned on about 120 h of Nigerian-language and code-switched speech with pause-normalised data (pause profile B);
-- in-context **anchor prompts** at the LM and the flow, so prosody and pause style follow the anchor voice (WAXAL Yoruba speaker 4);
-- a **flow-matching decoder fine-tuned on the anchor voice** (anchor SFT) for tone and pitch detail.
+**Code-switching TTS for Nigerian languages.** English, Nigerian Pidgin, Igbo, Yoruba and Hausa, mixed freely inside one sentence and spoken by one consistent voice, with no model swapping between languages.
 
-This repo is a **private research preview** for testing. It is not a production model.
+Dynamic Accent is a fine-tune of [FunAudioLLM/CosyVoice2-0.5B](https://huggingface.co/FunAudioLLM/CosyVoice2-0.5B) by Modular. It adds:
+- per-token **language embeddings**, a **speaker token** and an adversarial (gradient-reversal) speaker classifier on the LM, so the voice stays the same across a language switch (after Zhang et al., Google, 2019, *Learning to Speak Fluently in a Foreign Language*);
+- training on about 178 h of Nigerian-language and code-switched speech, including about 102 h of Igbo;
+- learned `[pause]`/`[breath]` events and a speaking-rate instruction (Phase 3);
+- a **flow-matching decoder fine-tuned on the anchor voice** for tone and pitch detail.
 
-Inference defaults to the **Phase 2 final behaviour** (`--pause_mode model --speed 1.0`): model-generated pauses. In a listening test the original clearly sounded more natural than prosody v2, so clause mode is kept only as an experimental option.
+It is a research release, not a production system. Read **Limitations** before you use it.
 
-## What is in the repo
+## Versions and voices
 
-| path | what it is |
-|---|---|
-| `checkpoints/p2_lm_B/best.pt` | Dynamic Accent LM (CosyVoice2 Qwen2 LM + language/speaker embeddings), pause profile B, 4,000 steps, best validation CE |
-| `checkpoints/p2_lm_B/spk2id.json` | speaker-token map (the anchor is `waxal_yor_4`) |
-| `checkpoints/p2_flow_sft/best.pt` | anchor-SFT flow-matching decoder (replaces the CosyVoice2 flow weights) |
-| `prompts/clip{0..3}_*.wav`, `prompts/prompts.json` | the 4 anchor prompt clips (24 kHz) with their transcripts and S3 speech tokens. The final system uses clip 0 |
-| `infer.py` | self-contained inference: the same path as the Phase 2 final samples |
-| `code/` | the original training and synthesis code (`train_b.py`, `train_b2.py`, `da/`) for reference |
-| `cosyvoice_requirements.txt` | CosyVoice's own requirements file, at the commit used |
-| `samples/` | example renders |
+Two model versions ship side by side. Pick one with `--version`, or render both with `--compare` and A/B before use.
+
+| version | LM | flow | decoding | what changed |
+|---|---|---|---|---|
+| `phase3` (**default**) | `checkpoints/phase3/p3_lm_full` (Phase 2 LM + 5,000 steps on a 178 h re-cut with ~102 h Igbo, `[pause]`/`[breath]` tags and a speed instruction, rate `normal`) | shared `checkpoints/phase2/p2_flow_sft` | one pass per paragraph, silence-aware repetition check, no audio post-processing | lower CER in Igbo/Hausa/Yoruba/code-switch/paragraphs, comma pauses at the anchor's length, clipped endings 63% → 11% |
+| `phase2` | `checkpoints/phase2/p2_lm_B` (~120 h, pause profile B) | `checkpoints/phase2/p2_flow_sft` | sentence by sentence, internal silences capped at 0.6 s | the Phase 2 final system |
+
+| voice (`--voice`) | who | notes |
+|---|---|---|
+| `anchor_male` (default) | WAXAL Yoruba speaker 4, male | the voice both versions were trained around |
+| `waxal_ibo_5_female` | WAXAL Igbo speaker 5, female, native Igbo | zero-shot (prompt clip, no fine-tune); best CER of the three on the female test subset |
+| `waxal_yor_5_female` | WAXAL Yoruba speaker 5, female, native Yoruba | zero-shot; strongest identity match, but its 3.8 s prompt has no pauses, so pauses are short and endings often clipped |
 
-Everything else comes from the stock base model **`FunAudioLLM/CosyVoice2-0.5B`** (Qwen tokenizer `CosyVoice-BlankEN`, `llm.pt`, `hift.pt` vocoder, `campplus.onnx`, `speech_tokenizer_v2.onnx`, `cosyvoice2.yaml`). Nothing in the base needs changing; `infer.py` downloads it on first run.
+**Recommended pairing** (author's listening verdict): **Phase 3 sounds best with the female voices; Phase 2 is fine for the male anchor.** `infer.py` prints a note when you pick another pairing. The female voices were evaluated with `phase3` only; they run with `phase2` but are untested there. Settings live in `settings.json`.
 
-## Setup (single GPU, e.g. Colab T4 or A100)
+## Setup (single GPU)
 
 ```bash
 # 1. CosyVoice code (the commit the model was built and tested on)
 git clone --recursive https://github.com/FunAudioLLM/CosyVoice.git
 cd CosyVoice && git checkout 074ca6dc9e80a2f424f1f74b48bdd7d3fea531cc && git submodule update --init --recursive && cd ..
 
-# 2. This repo (private: log in with a token that can read Modularcomputing)
-huggingface-cli login
+# 2. This repo
 huggingface-cli download Modularcomputing/dynamic-accent-preview --local-dir dynamic-accent-preview
 
 # 3. Python deps (Python 3.10; on Colab keep the preinstalled torch/torchaudio)
 pip install -r dynamic-accent-preview/requirements.txt
 ```
 
-The base model downloads to `dynamic-accent-preview/pretrained/CosyVoice2-0.5B` the first time you run `infer.py`. You can also point `--model_dir` (or `COSYVOICE2_DIR`) at an existing copy. Point `--cosyvoice_dir` (or `COSYVOICE_REPO`) at your CosyVoice clone if it is not inside the repo folder.
-
-The 0.5B model should fit comfortably on a 16 GB T4. This build was tested on an H100 only, not yet on Colab.
+The stock base model `FunAudioLLM/CosyVoice2-0.5B` (tokenizer, vocoder, speech tokenizer, speaker encoder) downloads to `pretrained/CosyVoice2-0.5B` on first run; nothing in it needs changing. Point `--model_dir` (or `COSYVOICE2_DIR`) at an existing copy, and `--cosyvoice_dir` (or `COSYVOICE_REPO`) at your CosyVoice clone. A 16 GB GPU is enough for the CLI; the Gradio app keeps both versions loaded. Tested on an H100 (bf16). `infer.py` also runs on CPU in fp32, which is slow. T4-class GPUs fall back to fp16, which has not been tested.
 
 ## Usage
 
 ```bash
 cd dynamic-accent-preview
-python infer.py --cosyvoice_dir ../CosyVoice \
-  --text "<eng>this is dynamic accent,</eng> <pcm>e dey sound nice abi?</pcm>" \
-  --seed 0 --out out.wav                     # defaults: --pause_mode model --speed 1.0
-# experimental prosody v2: add  --pause_mode clause --speed 1.05
-```
+# default: --version phase3 --voice anchor_male
+python infer.py --cosyvoice_dir ../CosyVoice --text "<eng>this is dynamic accent,</eng> <pcm>e dey sound nice abi?</pcm>" --out out.wav
 
-- **Language tags:** wrap each span in `<eng>`, `<pcm>`, `<yor>`, `<ibo>` or `<hau>`. Untagged text is read as English. Keep punctuation inside the span it belongs to.
-- **`--seed`:** the final system is a single sample at seed 0. Try seeds 1 and 2 (or more) and keep the take you like. Best-of-N picking lowered CER by about a third in evaluation.
-- **`--pause_mode clause`** (experimental, prosody v2; rejected in listening): the text is cut at `, ; : . ? !` and at language switches. Each chunk is generated separately and its edge silence is trimmed (energy VAD, 30 ms kept). Chunks are then joined with set gaps taken from the real anchor's pause statistics: 0.40 s ± 0.05 at commas and `;:`, 0.60 s at sentence ends, and 0.12 s at an unpunctuated language switch, with 10 ms fades. A one-word chunk is merged into the next chunk of the same language.
-- **`--speed`** (default 1.0): a pitch-preserving tempo factor (WSOLA via `audiotsm`) applied to the speech only, so the gaps stay as set. The allowed range is 1.0–1.2.
-- **`--pause_mode model --speed 1.0`** reproduces the Phase 2 final system: model-generated pauses, sentence-by-sentence generation, internal silences capped at 0.6 s, and 0.6 s between sentences. Both modes use anchor prompt clip 0 at the LM and flow (`--prompt_clip 0`).
-- **`--dtype`:** the LM ran under bf16 autocast. T4 has no bf16, so `auto` uses fp16 there. fp16 has not been tested yet; if a T4 run sounds wrong, compare against an A100 (bf16) run.
-- **Determinism:** generation is not bit-reproducible across runs on GPU. Sampling under bf16 can pick a different token on near-ties, so the same seed can give a slightly different take.
+# recommended pairing
+python infer.py --cosyvoice_dir ../CosyVoice --version phase3 --voice waxal_ibo_5_female \
+  --text "<ibo>Kedu ka ị mere?</ibo> <pcm>Abeg no vex, I go call you for evening.</pcm>" --out ibo_f.wav
 
-From Python:
+# A/B: the same text, seed and voice through both versions -> cmp_phase2.wav, cmp_phase3.wav
+python infer.py --cosyvoice_dir ../CosyVoice --compare --voice anchor_male --text "..." --out cmp.wav
 
-```python
-from infer import DynamicAccent
-tts = DynamicAccent("pretrained/CosyVoice2-0.5B", "../CosyVoice",
-                    "checkpoints/p2_lm_B/best.pt", "checkpoints/p2_flow_sft/best.pt", "prompts/prompts.json")
-wav = tts.say("<hau>Ina kwana, yaya aiki?</hau> <eng>The meeting has been moved to Friday morning.</eng>",
-              seed=0, pause_mode="model", speed=1.0)
-import soundfile as sf; sf.write("hau_eng.wav", wav, tts.sr)
+python infer.py --list_voices     # voices, versions, recommended pairing
 ```
 
-### Example code-switched inputs
+| flag | what it does |
+|---|---|
+| `--text` | the text: tagged (below) or plain with `--auto_tag` |
+| `--version phase2\|phase3` | model version (default `phase3`) |
+| `--voice anchor_male\|waxal_ibo_5_female\|waxal_yor_5_female` | voice (default `anchor_male`); old ids `anchor`, `waxal_ibo_5`, `waxal_yor_5` still work |
+| `--list_voices` | list voices, versions and the recommended pairing |
+| `--compare` | render with both versions → `<out>_phase2.wav`, `<out>_phase3.wav` |
+| `--rate default\|none\|very slow\|slow\|normal\|fast\|very fast` | Phase 3 speaking-rate instruction to the LM (default: the version's setting, `normal`). It is a learned instruction, not time-stretching. `phase2` ignores it |
+| `--auto_tag` | tag untagged mixed text automatically (see below), print the tagged text, then synthesise |
+| `--tag_only` | with `--auto_tag`: print the tagged text and stop, so you can edit it and pass it back as `--text` |
+| `--force_lang eng\|pcm\|ibo\|yor\|hau` | read the whole text in one language (tags are ignored) |
+| `--lang_map "omo=pcm,sha=pcm"` | per-word language overrides for the tagger |
+| `--lid none\|glotlid` | with `--auto_tag`: add [GlotLID v3](https://huggingface.co/cis-lmu/glotlid) clause scores (Apache-2.0 plus notices, 1.7 GB, `pip install fasttext`) |
+| `--seed N` | sampling seed (default 0). Try 2–3 seeds and keep the take you like |
+| `--out file.wav` | output path (24 kHz WAV) |
+| `--dtype auto\|bf16\|fp16\|fp32` | LM autocast dtype on GPU (samples: bf16). CPU always runs fp32 |
+| `--stitch` | `phase3`: sentence by sentence instead of one pass (not recommended) |
+| `--prompt_clip 0-3` | which anchor prompt clip (released systems: 0) |
+| `--model_dir`, `--cosyvoice_dir`, `--lm`, `--flow`, `--prompts`, `--voices` | paths and checkpoint overrides |
+| `--pause_mode clause`, `--speed` | `phase2` only: the experimental "prosody v2" (fixed gaps plus WSOLA speed-up). A listening test rejected it; the defaults (`model`, `1.0`) are the real Phase 2 system |
+
+Generation on GPU is not bit-reproducible: bf16 sampling can pick a different token on near-ties, so the same seed can give a slightly different take.
+
+### Controlling code-switching
+
+**1. Explicit tags (most control).** Wrap each span in `<eng>`, `<pcm>`, `<ibo>`, `<yor>` or `<hau>`. Keep punctuation inside the span it belongs to and put a space between spans. Untagged text is read as English.
 
 ```text
 <eng>this is dynamic accent,</eng> <pcm>e dey sound nice abi?</pcm>
+<pcm>omo,</pcm> <eng>I'm very excited to announce this new model,</eng> <ibo>ọ na-eme m obi ụtọ,</ibo> <pcm>I no go lie give you</pcm>
 <yor>Ẹ káàárọ̀ o, ṣé dáadáa ni?</yor> <eng>I will send you the report this afternoon.</eng>
-<ibo>Kedu ka ị mere?</ibo> <pcm>Abeg no vex, I go call you for evening.</pcm>
 <hau>Ina kwana, yaya aiki?</hau> <eng>The meeting has been moved to Friday morning.</eng>
 <ibo>Ndewo nụ, ụmụnna m.</ibo> <pcm>Una well done o, I know say the road no easy.</pcm> <eng>We have three things to discuss this evening.</eng>
-<hau>Barka da yamma, yaya kuke?</hau> <eng>The weather report says it will rain tomorrow morning.</eng> <pcm>Abeg no forget o, the rain fit heavy well well.</pcm>
+<eng>Please tell everyone that</eng> <yor>a ó pàdé ní ọ̀la.</yor>
 ```
 
-The Yoruba, Igbo and Hausa example lines are hand-written and have not been checked by native speakers.
+A switch can fall mid-sentence (`<eng>Please tell everyone that</eng> <yor>a ó pàdé ní ọ̀la.</yor>`). The tag sets the pronunciation and accent of the words inside it; the voice stays the same.
 
-## Samples
+**2. `--auto_tag` for untagged text.** `tagger.py` does word- and span-level language ID: Yoruba/Igbo/Hausa letters (ẹ ṣ / ị ụ ṅ / ɓ ɗ ƙ ƴ), small lexicons of frequent words, Pidgin markers ("dey", "abi", "omo", "wetin", "abeg", "una", "no go", "e don", "know say", pronoun + "go" + verb …) and clause context for ambiguous words ("na", "o", "ka", "ya"). It prints the tagged text so you can check and edit it:
+
+```bash
+python infer.py --auto_tag --tag_only --text "omo, I'm very excited to announce this new model, ọ na-eme m obi ụtọ, I no go lie give you"
+# [tagged] <pcm>omo,</pcm> <eng>I'm very excited to announce this new model,</eng> <ibo>ọ na-eme m obi ụtọ,</ibo> <pcm>I no go lie give you</pcm>
+python infer.py --auto_tag --text "..." --out out.wav            # tag, print, synthesise
+python tagger.py "Kedu ka ị mere? Abeg no vex, I go call you for evening."   # tagger on its own
+```
+
+Write Yoruba and Igbo with their diacritics where you can; text without diacritics is much harder to identify. See the tagger accuracy below.
+
+**3. `--force_lang`** reads everything in one language, e.g. `--force_lang ibo`.
+
+**4. `--lang_map`** fixes words the tagger gets wrong, e.g. `--lang_map "omo=pcm,sha=pcm,oga=pcm"`.
 
-- `samples/pcm_eng_dynamic_accent_seed{0,1,2}.mp3`: "this is dynamic accent, e dey sound nice abi?" (English then Pidgin), seeds 0, 1 and 2. These are Phase 2 final renders (`--pause_mode model --speed 1.0`).
-- `samples/prosody_v2/*.mp3`: the same custom sentence plus `para_ibo_pcm_eng`, `para_yor_eng`, `para_hau_eng_pcm`, `ibo_000` and `yor_000`, rendered with experimental prosody v2 (`--pause_mode clause --speed 1.05`, seed 0). The user preferred the original.
+### Gradio app (local)
 
-## Prosody v2 (inference-time pause and pace fix)
+```bash
+python app.py --cosyvoice_dir ../CosyVoice      # http://127.0.0.1:7860  (--share for a public link)
+```
+
+It has a text box, an auto-tag toggle with an editable tagged-text preview, word overrides, a force-language option, version, voice and rate dropdowns, a seed, and a **Compare** button that plays Phase 2 and Phase 3 side by side.
+
+### From Python
 
-**Speaking rate** is articulation rate: syllables per second of speech, with pauses of 0.15 s or more excluded. It uses the same syllable counter for every source. Reference speech is 150 clips per source. The system rate comes from the Phase 2 final renders of the eval set.
+```python
+import types, soundfile as sf, infer
+a = types.SimpleNamespace(model_dir="pretrained/CosyVoice2-0.5B", cosyvoice_dir="../CosyVoice", lm="", flow="",
+                          prompts="prompts/prompts.json", voices="prompts/voices.json", dtype="auto", stitch=False,
+                          rate="default", voice="waxal_ibo_5_female")
+tts = infer.build("phase3", a)
+wav = tts.say("<ibo>Kedu ka ị mere?</ibo> <eng>The meeting has been moved to Friday morning.</eng>", seed=0)
+sf.write("out.wav", wav, tts.sr)
+```
+
+## Samples
 
-| | anchor (yor) | yor | ibo | hau | pcm | eng |
-|---|---|---|---|---|---|---|
-| reference speech (syl/s) | 4.45 | 4.78 (WAXAL, other speakers) | 4.73 | 5.71 | 3.87 | 4.53 (FLEURS) |
-| Phase 2 final system (syl/s) | 4.20 (yor) | 4.20 | 4.50 | 3.98 | 3.63 | 4.77 |
-| reference ÷ system | 1.06 | 1.14 | 1.05 | 1.43 | 1.06 | 0.95 |
+Seed 0, rendered with `infer.py`. Inputs: `para_ibo_pcm_eng`, `para_yor_eng`, `para_hau_eng_pcm` (code-switched paragraphs), `ibo_000` (Igbo), `custom_dynacc` ("this is dynamic accent, e dey sound nice abi?") and `omo_announce` ("omo, I'm very excited to announce this new model, ọ na-eme m obi ụtọ, I no go lie give you").
 
-The median ratio is 1.06, so the default `--speed` is **1.05** (rounded to 0.05 and clamped to 1.0–1.2). Hausa is the outlier. Its vowel-based syllable count is the roughest of the five.
+| input | Phase 2, anchor_male | Phase 3, anchor_male | Phase 3, female |
+|---|---|---|---|
+| para_ibo_pcm_eng | [mp3](samples/phase3/phase2/para_ibo_pcm_eng.mp3) | [mp3](samples/phase3/phase3/para_ibo_pcm_eng.mp3) | [waxal_ibo_5_female](samples/phase3/phase3_female/para_ibo_pcm_eng__waxal_ibo_5.mp3) |
+| para_yor_eng | [mp3](samples/phase3/phase2/para_yor_eng.mp3) | [mp3](samples/phase3/phase3/para_yor_eng.mp3) | [waxal_yor_5_female](samples/phase3/phase3_female/para_yor_eng__waxal_yor_5.mp3) |
+| para_hau_eng_pcm | [mp3](samples/phase3/phase2/para_hau_eng_pcm.mp3) | [mp3](samples/phase3/phase3/para_hau_eng_pcm.mp3) | – |
+| ibo_000 | [mp3](samples/phase3/phase2/ibo_000.mp3) | [mp3](samples/phase3/phase3/ibo_000.mp3) | [waxal_ibo_5_female](samples/phase3/phase3_female/ibo_000__waxal_ibo_5.mp3) |
+| custom_dynacc | [mp3](samples/phase3/phase2/custom_dynacc.mp3) | [mp3](samples/phase3/phase3/custom_dynacc.mp3) | [waxal_yor_5_female](samples/phase3/phase3_female/custom_dynacc__waxal_yor_5.mp3) |
+| omo_announce | [mp3](samples/phase3/phase2/omo_announce.mp3) | [mp3](samples/phase3/phase3/omo_announce.mp3) | [waxal_ibo_5_female](samples/phase3/phase3_female/omo_announce__waxal_ibo_5.mp3), [waxal_yor_5_female](samples/phase3/phase3_female/omo_announce__waxal_yor_5.mp3) |
 
-**Before and after** on the six sample items (custom sentence, 3 paragraphs, ibo_000, yor_000), seed 0. Pauses are measured at forced-aligned word boundaries with the Phase 2 report harness. The anchor row is real WAXAL recordings. With only 6 items, read these as a direction, not a result.
+Older renders: `samples/pcm_eng_dynamic_accent_seed{0,1,2}.mp3` (Phase 2, three seeds) and `samples/prosody_v2/` (the rejected prosody-v2 post-processing, kept for reference).
 
-| | comma pause (s) | sentence pause (s) | unpunctuated pause rate (>0.25 s) | articulation rate (syl/s) | para CER | para CER (no tones) | all-6 CER | SIM |
+## Evaluation: Phase 2 vs Phase 3
+
+183-item eval set (30 per language, 10 code-switched, 3 hand-written + 20 native long-form paragraphs), anchor voice, seed 0. CER is ASR round-trip: MMS-1b-all for Igbo/Hausa/Yoruba, Whisper-large-v3-turbo for English and Pidgin, the lower of the two for code-switched items and paragraphs. "No tones" (in brackets) strips diacritics. SIM is ECAPA cosine to the anchor (ceiling about 0.84 on real held-out anchor clips). The Pidgin judge is weak: real Pidgin recordings score about 0.43.
+
+| system | ibo | hau | yor | pcm | eng | CS | para (hand / native) | SIM |
 |---|---|---|---|---|---|---|---|---|
-| real anchor | 0.417 | 0.543 | 0.137 | 4.45 | – | – | – | ceiling ≈0.84 |
-| before: Phase 2 final (`model`, 1.0) | 0.339 | 0.488 | 0.057 | 4.26 | 0.175 | 0.124 | 0.175 | 0.814 |
-| clause, speed 1.0 | 0.339 | 0.526 | 0.095 | 3.95 | 0.179 | 0.123 | 0.166 | 0.775 |
-| after: clause, speed 1.05 (experimental) | 0.275 (≈0.37 *) | 0.550 | 0.086 | 4.18 | 0.193 | 0.143 | 0.207 | 0.750 |
-| clause, speed 1.10 | 0.315 | 0.550 | 0.067 | 4.37 | 0.244 | 0.194 | 0.224 | 0.748 |
-| model pauses, speed 1.05 | 0.235 | 0.441 | 0.029 | 4.35 | 0.189 | 0.149 | 0.188 | 0.788 |
+| `phase2` as shipped (sentence stitching) | 0.215 (0.171) | 0.194 | 0.234 (0.149) | 0.196 | 0.035 | 0.163 | 0.192 / 0.154 | 0.785 |
+| Phase 2 LM, Phase 3 decoding | 0.243 (0.189) | 0.169 | 0.275 (0.175) | 0.195 | 0.046 | 0.185 | 0.195 / 0.158 | 0.787 |
+| **`phase3`** | **0.219 (0.173)** | **0.151** | **0.233 (0.145)** | 0.202 | 0.043 | **0.151** | **0.170** / 0.155 | **0.788** |
 
-\* At two commas that follow the one-letter Pidgin particle "o," the aligner moves the inserted gap to the boundary before "o". This reads as a 0.0 s comma plus one unpunctuated pause. The other comma gaps in clause mode measure 0.37–0.39 s, against the 0.40 s target.
+Same decoding, Phase 3 LM vs Phase 2 LM: lower CER in every language except Pidgin, same speaker similarity.
 
-- **Pauses:** clause mode moves comma and sentence pauses and the phrase-break rate toward the anchor's numbers. Sentence pauses go from 0.49 to 0.55 s against the anchor's 0.54 s.
-- **Pace:** generating short chunks one at a time makes speech about 7% slower (4.26 to 3.95 syl/s). So 1.05 only brings the rate back to about where the final system was, and it is still below the anchor.
-- **Cost:** WSOLA speed-up and chunking each lower ECAPA similarity (0.81 to 0.75–0.79). At 1.10, paragraph CER rises clearly (0.175 to 0.244). At 1.05 the rise is small (0.175 to 0.193), and mixed across paragraphs: Hausa got better, Igbo and Yoruba got worse. A blind listening check should decide between 1.0 and 1.05.
+| system | comma pause s | sentence pause s | pauses / s | pause-length W1 vs anchor | trailing silence s | clipped endings (< 0.1 s) | articulation syl/s |
+|---|---|---|---|---|---|---|---|
+| real anchor recordings | 0.417 | 0.543 | 0.624 | 0 | 0.16 | 15% | 4.43 |
+| Phase 2 LM, Phase 3 decoding | 0.403 | 0.389 | 0.329 | 0.194 | 0.08 | 63% | 4.36 |
+| **`phase3`** | **0.422** | **0.453** | 0.319 | **0.161** | **0.17** | **11%** | 4.44 |
 
-## Evaluation (Phase 2 report)
+Female voices (29-item subset, `phase3`):
 
-CER is ASR round-trip: MMS-1b-all for ibo/hau/yor, Whisper-large-v3-turbo for eng and pcm, and min(MMS, Whisper) for code-switched (CS) items and paragraphs. "No tones" strips the diacritics. SIM is ECAPA cosine to the anchor centroid; the ceiling is about 0.84 on real held-out anchor clips. Pidgin CER is limited by the judge: real Pidgin recordings score 0.43. Full eval set: 30 items per language + 9 CS sentences + 3 paragraphs = 162 items, single sample (seed 0, anchor prompt).
+| voice | CER | no-tone CER | SIM to the speaker's own real clips | clipped endings |
+|---|---|---|---|---|
+| `anchor_male` (same items) | 0.182 | 0.141 | 0.12–0.13 (it is a different speaker) | 17% |
+| `waxal_ibo_5_female` | **0.149** | **0.101** | 0.64 | 21% |
+| `waxal_yor_5_female` | 0.171 | 0.134 | **0.76** | 62% |
 
-| system | ibo CER | hau CER | yor CER (no tones) | pcm CER | eng CER | CS CER | para CER | SIM to anchor (mean of 5 langs) | SIM spread | CS half-vs-half SIM |
-|---|---|---|---|---|---|---|---|---|---|---|
-| Phase 1 Arm B (smoke) | 0.212 | 0.244 | 0.407 (0.294) | 0.412 | 0.035 | 0.201 | 0.195 | 0.804 | 0.030 | 0.649 |
-| p2_lm (profile A) + SFT flow | 0.199 | 0.183 | 0.223 (0.140) | 0.190 | 0.038 | 0.184 | 0.176 | 0.769 | 0.069 | 0.634 |
-| **p2_lm_B (profile B) + SFT flow (this release)** | 0.188 | 0.188 | 0.224 (0.137) | 0.174 | 0.034 | 0.167 | 0.173 | 0.782 | 0.039 | 0.642 |
+Naturalness has not been rated by native speakers (no MOS or blind A/B yet). The numbers above are automatic proxies.
 
-Pause and pitch (energy VAD on forced-aligned word boundaries):
+### Auto-tagger accuracy
 
-| system | int. pause p90 (s) | pauses / s | spurious pause rate | comma pause (s) | pause-length W1 vs anchor | para: sentence pause (s) | para: pause p90 (s) | F0 std (st) | F0 range (st) | yor tone-direction acc (pairs) |
-|---|---|---|---|---|---|---|---|---|---|---|
-| **real anchor recordings** (WAXAL val+test, raw) | 1.071 | 0.624 | 0.137 | 0.417 | 0 | 0.543 | – | 3.152 | 10.173 | 0.750 (2990) |
-| Phase 1 Arm B (smoke) | 1.334 | 0.584 | 0.044 | 0.245 | 0.147 | 0.315 | 1.035 | 3.058 | 9.999 | 0.777 (251) |
-| p2_lm (profile A) + SFT flow | 0.414 | 0.328 | 0.008 | 0.263 | 0.330 | 0.509 | 0.547 | 3.038 | 10.069 | 0.752 (254) |
-| **p2_lm_B (profile B) + SFT flow (this release)** | 0.460 | 0.288 | 0.087 | 0.251 | 0.262 | 0.488 | 0.531 | 3.057 | 10.134 | 0.749 (255) |
+Word-level accuracy of `tagger.py` against hand-tagged text (no GlotLID / with `--lid glotlid`):
 
-Zero-shot baselines on the 62-item stage subset (for context): CosyVoice2 zero-shot ibo/hau/yor CER 0.689 / 0.281 / 0.852, and CosyVoice3 zero-shot 0.629 / 0.197 / 0.717, against 0.173 / 0.220 / 0.226 for Phase 2 (p2_lm + SFT flow).
+| test set | words | rule-based | + GlotLID | exact sentences (rule-based) |
+|---|---|---|---|---|
+| the 6 sample inputs above | 145 | 1.000 | 1.000 | 6/6 |
+| 10 new mixed sentences, written before the lexicons | 113 | 0.991 | 0.991 | 9/10 |
+| 12 code-switched eval items (overlaps the samples) | 206 | 0.995 | 1.000 | 11/12 |
+| 150 single-language eval sentences (real WAXAL/FLEURS text, never looked at while building the lexicons) | 1,697 | **0.898** | **0.915** | 115/150 |
 
-## Known limitations
+On the held-out single-language set, per language (rule-based / + GlotLID): English 0.99 / 1.00, Igbo 1.00 / 1.00, Yoruba 1.00 / 1.00, Hausa 0.89 / 0.94, **Pidgin 0.63 / 0.65**. Pidgin written without markers looks like English and gets tagged `eng`. The mixed test sentences were written and gold-tagged by the tagger's author, not native speakers, so their scores are optimistic. Check the printed tags, and use explicit tags or `--lang_map` when it matters.
 
-- In `model` mode, comma pauses are short (0.25 s on the full eval set, against the anchor's 0.42 s). `clause` mode sets them directly (see Prosody v2).
-- Speaker similarity is 0.78 against Phase 1's 0.80. English and Pidgin drift the most from the anchor.
+## Limitations
+
+- **Pace**: speech is still 13–28% slower than native speakers (Igbo 4.60 vs 5.30 syl/s, Hausa 3.77 vs 5.26 syl/s), and there are about half as many pauses per second as in real speech. It can sound careful rather than fully native.
+- **Pidgin is the weakest language**: only about 5 h of training data, no reliable ASR judge, and the auto-tagger's Pidgin/English split is its weakest point (~0.63 word accuracy on plain Pidgin).
+- **No laughter or other expressive events**: the training data has effectively no laughter, so tags like `[laughter]` do nothing useful. `[pause]`/`[breath]` were learned from data, but there is no reliable user control over them yet.
 - English after a Nigerian-language span keeps a strong local accent.
-- Pidgin has only about 6 h of training data and no reliable ASR judge.
-- Yoruba tone quality needs human listening; the automatic tone metric is saturated.
-- Naturalness has not been rated by native speakers yet (no MOS / A/B).
+- Yoruba tone quality needs native listening; the automatic tone metric is saturated.
+- Female voices are zero-shot; `waxal_yor_5_female` copies its short prompt (short pauses, clipped endings).
+- Not evaluated for numbers, abbreviations or URLs; write numbers as words.
+- Do not use it to impersonate real people. The voices are WAXAL Media Trust speakers, released under CC-BY-SA-4.0 for speech technology.
 
-## Training data
+## What is in the repo
 
-The training data includes CC-BY-SA sources. Datasets used:
-- **WAXAL** (Yoruba, Igbo, Hausa, Nigerian Pidgin; includes the anchor voice, Yoruba speaker 4), CC-BY-SA-4.0
-- **BibleTTS** (Yoruba and Hausa), CC-BY-SA
-- **FLEURS** (English, Hausa, Igbo, Yoruba)
-- **OpenSLR SLR70** (Nigerian English)
-- **OpenSLR SLR86** (Yoruba)
-- 4,000 code-switched pairs built from the sets above
+| path | what it is | licence |
+|---|---|---|
+| `checkpoints/phase3/p3_lm_full/` | Phase 3 LM (`best.pt`, val step 4,500), `config.json` (rate, sampler, one-pass, points at the shared Phase 2 flow), `spk2id.json`, `args.json`, `log.jsonl` | CC-BY-SA-4.0 |
+| `checkpoints/phase2/p2_lm_B/` | Phase 2 LM (`best.pt`) and `spk2id.json` | CC-BY-SA-4.0 |
+| `checkpoints/phase2/p2_flow_sft/best.pt` | anchor-SFT flow-matching decoder, used by both versions | CC-BY-SA-4.0 |
+| `prompts/` | anchor prompt clips (`clip0..3`, `prompts.json` with transcripts and S3 tokens) and female voice prompts (`voice_*.wav`, `voices.json`), all from WAXAL | CC-BY-SA-4.0 |
+| `samples/` | example renders | CC-BY-SA-4.0 |
+| `settings.json` | versions, voices, recommended pairing | Apache-2.0 |
+| `infer.py`, `tagger.py`, `app.py` | inference CLI, auto-tagger, Gradio app | Apache-2.0 |
+| `code/` | the original training and synthesis code, for reference | Apache-2.0 |
+| `requirements.txt`, `cosyvoice_requirements.txt` | dependencies (`cosyvoice_requirements.txt` is CosyVoice's own file at the commit used) | – |
+| `assets/` | banner | – |
+
+## Training data and attribution
+
+The model weights are trained on, and the voice prompts are taken from, the datasets below. Several are CC-BY-SA-4.0, so the weights, voice prompts and samples are released under **CC-BY-SA-4.0** (share-alike).
+
+| dataset | used for | licence | attribution |
+|---|---|---|---|
+| [WAXAL](https://huggingface.co/datasets/google/WaxalNLP), Media Trust TTS subsets (Igbo, Yoruba, Hausa, Nigerian Pidgin) | main training data; the **anchor voice** (Yoruba speaker 4) and the **female prompt voices** (Igbo speaker 5, Yoruba speaker 5) | CC-BY-SA-4.0 (as listed on the dataset card for the Media Trust subsets) | Google and Media Trust, WAXAL (arXiv:2602.02734) |
+| [Open.Bible Igbo](https://huggingface.co/datasets/AfriSpeech/open-bible-speech-african) (Biblica® Open Igbo Contemporary Bible™, Audio Edition) | Igbo (Phase 3) | CC-BY-SA-4.0 | Biblica, Inc. via open.bible; Biblica® and Open Igbo Contemporary Bible™ are trademarks of Biblica, Inc. |
+| [NaijaS2ST](https://huggingface.co/datasets/McGill-NLP/NaijaS2ST) (Igbo) | Igbo, many native speakers (Phase 3) | CC-BY-4.0 | McGill-NLP, NaijaS2ST (arXiv:2604.16287) |
+| [IgboAPI](https://huggingface.co/datasets/nkowaokwu/ibo-dict) (ibo-dict, ibo-dict-expansion) | Igbo sentences (Phase 3) | CC-BY-4.0 | Nkọwa okwu / IgboAPI contributors |
+| [IgboSynCorp](https://doi.org/10.7910/dvn/yb9fwk) | natural Igbo speech (Phase 3) | CC0-1.0 | University of Ibadan and Afe Babalola University (Lacuna Fund) |
+| [FLEURS](https://huggingface.co/datasets/google/fleurs) (English, Hausa, Igbo, Yoruba) | multi-speaker data | CC-BY-4.0 | Google, FLEURS (Conneau et al., 2022) |
+| [BibleTTS](https://masakhane-io.github.io/bibleTTS/) (Yoruba, Hausa) | Yoruba and Hausa | CC-BY-SA-4.0 | Meyer et al., BibleTTS (2022); Biblica / open.bible audio |
+| [OpenSLR SLR70](https://openslr.org/70/) (Nigerian English) | English | CC-BY-SA-4.0 | Google, crowdsourced Nigerian English |
+| [OpenSLR SLR86](https://openslr.org/86/) (Yoruba) | Yoruba | CC-BY-SA-4.0 | Google, crowdsourced Yoruba |
+| code-switched pairs | about 9 h of synthetic code-switch training pairs | follows the source clips (CC-BY-SA-4.0 / CC-BY-4.0) | built by concatenating clips from the sets above, following the Google 2019 code-switching recipe (Zhang et al., 2019); no separate Google dataset |
+| base model [FunAudioLLM/CosyVoice2-0.5B](https://huggingface.co/FunAudioLLM/CosyVoice2-0.5B) | LM initialisation, flow, vocoder, tokenizers | Apache-2.0 | FunAudioLLM / Alibaba Tongyi Lab, CosyVoice 2 (Du et al., 2024) |
+| optional [GlotLID](https://huggingface.co/cis-lmu/glotlid) (not bundled) | `--lid glotlid` in the tagger | Apache-2.0 plus notices | Kargaran et al., GlotLID (EMNLP 2023) |
+
+Excluded on licence grounds: WaZoBiaSpeech (bans TTS use), NaijaVoices (non-commercial), and non-commercial TTS weights.
 
 ## Licence
 
-Apache-2.0 (see `LICENSE`). The base model, CosyVoice2-0.5B, is Apache-2.0.
+- **Model weights, voice prompts and samples** (`checkpoints/`, `prompts/`, `samples/`): **[CC-BY-SA-4.0](LICENSE)**. You may use them commercially. You must credit Modular and the datasets listed above, and anything you derive from the weights (fine-tunes, merges) must be shared under CC-BY-SA-4.0 too.
+- **Code** (`infer.py`, `tagger.py`, `app.py`, `code/`, `settings.json`): **[Apache-2.0](LICENSE-CODE)**.
+- The base model, CosyVoice2-0.5B, is Apache-2.0; its notices still apply to the parts it contributes.
+
+## Citation
+
+```bibtex
+@misc{dynamicaccent2026,
+  title  = {Dynamic Accent: Code-switching TTS for Nigerian languages},
+  author = {Modular},
+  year   = {2026},
+  url    = {https://huggingface.co/Modularcomputing/dynamic-accent-preview}
+}
+```

app.py

@@ -0,0 +1,140 @@
+#!/usr/bin/env python
+"""Dynamic Accent: local Gradio demo (single GPU).
+
+    pip install -r requirements.txt            # includes gradio
+    python app.py --cosyvoice_dir ../CosyVoice  # then open http://127.0.0.1:7860
+
+Text box -> optional auto-tag (editable preview of the tagged text) -> version + voice + rate -> Synthesize, or
+Compare to hear Phase 2 and Phase 3 on the same text, seed and voice. Both versions stay loaded (one shared CosyVoice2
+base; each version has its own LM copy), so a GPU with >= 16 GB is recommended. CPU works but is slow.
+"""
+import argparse, json, os, sys, types
+
+HERE = os.path.dirname(os.path.abspath(__file__))
+sys.path.insert(0, HERE)
+import infer  # noqa: E402
+from tagger import Tagger, parse_map, strip_tags, TAGGED, LANGS  # noqa: E402
+
+EXAMPLES = [
+    "Omo, I'm very excited to announce this new model, ọ na-eme m obi ụtọ, I no go lie give you",
+    "Ẹ káàbọ̀ sí ìpàdé wa òní. Today we will talk about the new school calendar.",
+    "Barka da yamma, yaya kuke? The weather report says it will rain tomorrow morning.",
+    "Kedu ka ị mere? Abeg no vex, I go call you for evening.",
+]
+
+
+class Engine:
+    def __init__(self, a):
+        self.a = a
+        self.models = {}
+        self.cosy = self.tok = None
+        self.taggers = {}
+
+    def get(self, version, voice):
+        if version not in self.models:
+            ns = types.SimpleNamespace(**vars(self.a))
+            ns.voice = voice
+            S = infer.build(version, ns, self.cosy, self.tok, copy_llm=self.cosy is not None)
+            self.cosy, self.tok = S.cosy, S.tok
+            S._rate0 = S.rate
+            self.models[version] = S
+        S = self.models[version]
+        if getattr(S, "_voice_name", None) != voice:
+            S.set_voice(voice, self.a.voices); S._voice_name = voice
+        return S
+
+    def tag(self, text, lang_map, use_lid):
+        key = (lang_map, use_lid)
+        if key not in self.taggers:
+            self.taggers[key] = Tagger("glotlid" if use_lid else None, parse_map(lang_map))
+        return self.taggers[key].tag(strip_tags(text))
+
+    def say(self, text, version, voice, rate, seed):
+        S = self.get(version, voice)
+        S.rate = S._rate0 if rate == "default" else (infer.RATES.index(rate) if rate in infer.RATES else 0)
+        w = S.say(text, seed=int(seed))
+        return S.sr, w
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--cosyvoice_dir", default=os.environ.get("COSYVOICE_REPO", os.path.join(HERE, "CosyVoice")))
+    ap.add_argument("--model_dir", default=os.environ.get("COSYVOICE2_DIR", os.path.join(HERE, "pretrained", "CosyVoice2-0.5B")))
+    ap.add_argument("--host", default="127.0.0.1")
+    ap.add_argument("--port", type=int, default=7860)
+    ap.add_argument("--share", action="store_true")
+    a = ap.parse_args()
+    for k, v in dict(lm="", flow="", prompts=os.path.join(HERE, "prompts", "prompts.json"),
+                     voices=os.path.join(HERE, "prompts", "voices.json"), dtype="auto", stitch=False, rate="default").items():
+        setattr(a, k, v)
+    if not os.path.isdir(a.model_dir):
+        from huggingface_hub import snapshot_download
+        snapshot_download("FunAudioLLM/CosyVoice2-0.5B", local_dir=a.model_dir)
+
+    import gradio as gr
+    E = Engine(a)
+    S = infer.SETTINGS
+    voices = S.get("voices", ["anchor_male", "waxal_ibo_5_female", "waxal_yor_5_female"])
+    rec = S.get("recommended", {})
+    rates = ["default"] + S.get("rates", ["none", "very slow", "slow", "normal", "fast", "very fast"])
+
+    def prepare(text, auto, lang_map, use_lid, force):
+        if not text.strip():
+            return ""
+        if force and force != "off":
+            return f"<{force}>{strip_tags(text)}</{force}>"
+        if auto and not TAGGED.search(text):
+            return E.tag(text, lang_map or "", bool(use_lid))
+        return text
+
+    def hint(version, voice):
+        r = rec.get(version, [])
+        return "" if voice in r else f"Tip: {version} is recommended with {', '.join(r)}. {rec.get('note', '')}"
+
+    def synth(text, tagged, auto, lang_map, use_lid, force, version, voice, rate, seed):
+        t = tagged.strip() if tagged.strip() else prepare(text, auto, lang_map, use_lid, force)
+        sr, w = E.say(t, version, voice, rate, seed)
+        return t, (sr, w), hint(version, voice)
+
+    def compare(text, tagged, auto, lang_map, use_lid, force, voice, rate, seed):
+        t = tagged.strip() if tagged.strip() else prepare(text, auto, lang_map, use_lid, force)
+        a2 = E.say(t, "phase2", voice, rate, seed)
+        a3 = E.say(t, "phase3", voice, rate, seed)
+        return t, a2, a3
+
+    with gr.Blocks(title="Dynamic Accent: code-switching TTS") as demo:
+        gr.Markdown("## Dynamic Accent: code-switching TTS\nEnglish, Nigerian Pidgin, Igbo, Yoruba and Hausa, mixed in one "
+                    "sentence. Tag spans with `<eng> <pcm> <ibo> <yor> <hau>`, or turn on auto-tag and edit its preview.")
+        text = gr.Textbox(label="Text", lines=3, value=EXAMPLES[0])
+        with gr.Row():
+            auto = gr.Checkbox(label="Auto-tag languages", value=True)
+            use_lid = gr.Checkbox(label="Use GlotLID (1.7 GB download, needs fasttext)", value=False)
+            force = gr.Dropdown(["off"] + LANGS, value="off", label="Force one language")
+            lang_map = gr.Textbox(label="Word overrides (e.g. omo=pcm,sha=pcm)", value="")
+        tag_btn = gr.Button("Preview tags")
+        tagged = gr.Textbox(label="Tagged text (edit before synthesis; leave empty to tag automatically)", lines=3)
+        with gr.Row():
+            version = gr.Dropdown(sorted(infer.VERSIONS), value=S.get("default_version", "phase3"), label="Version")
+            voice = gr.Dropdown(voices, value=S.get("default_voice", "anchor_male"), label="Voice")
+            rate = gr.Dropdown(rates, value="default", label="Rate (phase3)")
+            seed = gr.Number(value=0, precision=0, label="Seed")
+        tip = gr.Markdown()
+        with gr.Row():
+            go = gr.Button("Synthesize", variant="primary")
+            cmp_btn = gr.Button("Compare Phase 2 vs Phase 3")
+        out = gr.Audio(label="Output", type="numpy")
+        with gr.Row():
+            out2 = gr.Audio(label="Phase 2", type="numpy")
+            out3 = gr.Audio(label="Phase 3", type="numpy")
+        gr.Examples(EXAMPLES, inputs=text)
+
+        tag_btn.click(prepare, [text, auto, lang_map, use_lid, force], tagged)
+        version.change(hint, [version, voice], tip)
+        voice.change(hint, [version, voice], tip)
+        go.click(synth, [text, tagged, auto, lang_map, use_lid, force, version, voice, rate, seed], [tagged, out, tip])
+        cmp_btn.click(compare, [text, tagged, auto, lang_map, use_lid, force, voice, rate, seed], [tagged, out2, out3])
+    demo.queue(max_size=8).launch(server_name=a.host, server_port=a.port, share=a.share)
+
+
+if __name__ == "__main__":
+    main()

checkpoints/phase2/p2_lm_B/spk2id.json

@@ -0,0 +1 @@
+{"bible_hau": 0, "bible_yor": 1, "fleurs_eng_FEMALE_c000": 2, "fleurs_eng_FEMALE_c001": 3, "fleurs_eng_FEMALE_c002": 4, "fleurs_eng_FEMALE_c003": 5, "fleurs_eng_FEMALE_c004": 6, "fleurs_eng_FEMALE_c005": 7, "fleurs_eng_FEMALE_c006": 8, "fleurs_eng_FEMALE_c007": 9, "fleurs_eng_FEMALE_c008": 10, "fleurs_eng_FEMALE_c009": 11, "fleurs_eng_FEMALE_c010": 12, "fleurs_eng_FEMALE_c012": 13, "fleurs_eng_MALE_c000": 14, "fleurs_eng_MALE_c001": 15, "fleurs_eng_MALE_c002": 16, "fleurs_eng_MALE_c003": 17, "fleurs_eng_MALE_c004": 18, "fleurs_eng_MALE_c005": 19, "fleurs_eng_MALE_c006": 20, "fleurs_eng_MALE_c007": 21, "fleurs_hau_FEMALE_c000": 22, "fleurs_hau_FEMALE_c001": 23, "fleurs_hau_FEMALE_c003": 24, "fleurs_hau_FEMALE_misc": 25, "fleurs_hau_MALE_c000": 26, "fleurs_hau_MALE_c002": 27, "fleurs_hau_MALE_c003": 28, "fleurs_hau_MALE_c004": 29, "fleurs_hau_MALE_misc": 30, "fleurs_ibo_FEMALE_c000": 31, "fleurs_ibo_FEMALE_c001": 32, "fleurs_ibo_FEMALE_c002": 33, "fleurs_ibo_FEMALE_c003": 34, "fleurs_ibo_FEMALE_c004": 35, "fleurs_ibo_FEMALE_misc": 36, "fleurs_ibo_MALE_c000": 37, "fleurs_ibo_MALE_c001": 38, "fleurs_ibo_MALE_c002": 39, "fleurs_ibo_MALE_c004": 40, "fleurs_ibo_MALE_c005": 41, "fleurs_ibo_MALE_misc": 42, "fleurs_yor_FEMALE_c000": 43, "fleurs_yor_FEMALE_c001": 44, "fleurs_yor_FEMALE_c002": 45, "fleurs_yor_MALE_c000": 46, "fleurs_yor_MALE_c001": 47, "fleurs_yor_MALE_misc": 48, "slr70_ngf_00295": 49, "slr70_ngf_00610": 50, "slr70_ngf_01208": 51, "slr70_ngf_01523": 52, "slr70_ngf_02121": 53, "slr70_ngf_02436": 54, "slr70_ngf_02484": 55, "slr70_ngf_03349": 56, "slr70_ngf_03397": 57, "slr70_ngf_04310": 58, "slr70_ngf_05223": 59, "slr70_ngf_06136": 60, "slr70_ngf_07049": 61, "slr70_ngf_07508": 62, "slr70_ngf_08421": 63, "slr70_ngf_09334": 64, "slr70_ngf_12484": 65, "slr70_ngf_13397": 66, "slr70_ngf_15223": 67, "slr70_ngm_01523": 68, "slr70_ngm_02436": 69, "slr70_ngm_02484": 70, "slr70_ngm_03349": 71, "slr70_ngm_03397": 72, "slr70_ngm_07049": 73, "slr70_ngm_07508": 74, "slr70_ngm_08784": 75, "slr70_ngm_09334": 76, "slr70_ngm_09697": 77, "slr70_ngm_12484": 78, "slr70_ngm_14310": 79, "slr86_yof_00295": 80, "slr86_yof_00610": 81, "slr86_yof_01208": 82, "slr86_yof_02121": 83, "slr86_yof_02436": 84, "slr86_yof_02484": 85, "slr86_yof_03034": 86, "slr86_yof_03349": 87, "slr86_yof_03397": 88, "slr86_yof_04310": 89, "slr86_yof_05223": 90, "slr86_yof_06136": 91, "slr86_yof_07049": 92, "slr86_yof_07505": 93, "slr86_yof_07508": 94, "slr86_yof_08421": 95, "slr86_yof_08784": 96, "slr86_yof_09334": 97, "slr86_yof_09697": 98, "slr86_yom_00295": 99, "slr86_yom_00610": 100, "slr86_yom_01208": 101, "slr86_yom_01523": 102, "slr86_yom_02121": 103, "slr86_yom_02436": 104, "slr86_yom_02484": 105, "slr86_yom_03034": 106, "slr86_yom_03397": 107, "slr86_yom_04310": 108, "slr86_yom_06136": 109, "slr86_yom_07049": 110, "slr86_yom_07505": 111, "slr86_yom_07508": 112, "slr86_yom_08421": 113, "slr86_yom_08784": 114, "slr86_yom_09334": 115, "waxal_hau_1": 116, "waxal_hau_2": 117, "waxal_hau_3": 118, "waxal_hau_4": 119, "waxal_hau_5": 120, "waxal_hau_6": 121, "waxal_hau_7": 122, "waxal_hau_8": 123, "waxal_ibo_1": 124, "waxal_ibo_2": 125, "waxal_ibo_3": 126, "waxal_ibo_4": 127, "waxal_ibo_5": 128, "waxal_ibo_6": 129, "waxal_ibo_7": 130, "waxal_ibo_8": 131, "waxal_pcm_F1": 132, "waxal_pcm_F2": 133, "waxal_pcm_F3": 134, "waxal_pcm_F4": 135, "waxal_pcm_M1": 136, "waxal_pcm_M2": 137, "waxal_pcm_M3": 138, "waxal_pcm_M4": 139, "waxal_yor_1": 140, "waxal_yor_2": 141, "waxal_yor_3": 142, "waxal_yor_4": 143, "waxal_yor_5": 144, "waxal_yor_6": 145, "waxal_yor_7": 146, "waxal_yor_8": 147}

checkpoints/phase3/p3_lm_full/args.json

@@ -0,0 +1 @@
+{"cmd": "lm", "arm": "b", "out": "/home/dev/runs/p3_lm_full", "init": "/home/dev/runs/p2_lm_B/best.pt", "flow_init": "/home/dev/runs/p2_flow_sft/best.pt", "minutes": 60.0, "max_steps": 5000, "lr": 1e-05, "lr_new": 0.0005, "warmup": 200, "tokens": 14000, "max_tok": 750, "adv_w": 0.02, "p_prompt": 0.5, "tag_drop": 0.5, "rate_drop": 0.5, "anchor_w": 1.0, "val_every": 500, "patience": 4, "hours_cap": 0, "quad": 600.0}

checkpoints/phase3/p3_lm_full/best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1fa9986713767e80b3d6a6b72cbb3d4e182e37179554d34bdb4dc7a86cd52b2c
+size 2046917388

checkpoints/phase3/p3_lm_full/config.json

@@ -0,0 +1,7 @@
+{
+ "rate": "normal",
+ "flow": "checkpoints/phase2/p2_flow_sft/best.pt",
+ "sampler": "silexcap",
+ "one_pass": true,
+ "note": "Phase 3 LM (p3_lm_full, best val step 4500). Uses the shared Phase 2 anchor-SFT flow; a Phase 3 flow SFT gave no measurable gain and is not shipped."
+}

checkpoints/phase3/p3_lm_full/log.jsonl

@@ -0,0 +1,211 @@
+{"step": 0, "val": {"np": 3.563712294792896, "p": 3.571646510972262, "anchor_p": 3.2171030044555664, "mean": 3.567679402882579}}
+{"step": 25, "ce": 3.7304, "acc": 0.187, "adv": 8.175, "adv_acc": 0.129, "gn": 2.12, "B": 18, "lr": 1.25e-06, "skips": 0, "oom": 0, "min": 0.24}
+{"step": 50, "ce": 4.0527, "acc": 0.1759, "adv": 7.922, "adv_acc": 0.0, "gn": 3.5, "B": 10, "lr": 2.499e-06, "skips": 0, "oom": 0, "min": 0.36}
+{"step": 75, "ce": 4.1871, "acc": 0.1463, "adv": 8.59, "adv_acc": 0.009, "gn": 2.74, "B": 53, "lr": 3.748e-06, "skips": 0, "oom": 0, "min": 0.49}
+{"step": 100, "ce": 3.4694, "acc": 0.2184, "adv": 4.838, "adv_acc": 0.0, "gn": 2.66, "B": 6, "lr": 4.996e-06, "skips": 1, "oom": 0, "min": 0.61}
+{"step": 125, "ce": 3.7797, "acc": 0.1892, "adv": 5.106, "adv_acc": 0.029, "gn": 2.78, "B": 29, "lr": 6.241e-06, "skips": 1, "oom": 0, "min": 0.73}
+{"step": 150, "ce": 3.6674, "acc": 0.2085, "adv": 4.72, "adv_acc": 0.001, "gn": 2.17, "B": 14, "lr": 7.485e-06, "skips": 2, "oom": 0, "min": 0.85}
+{"step": 175, "ce": 3.8224, "acc": 0.178, "adv": 6.07, "adv_acc": 0.007, "gn": 2.08, "B": 34, "lr": 8.726e-06, "skips": 2, "oom": 0, "min": 0.97}
+{"step": 200, "ce": 3.4966, "acc": 0.231, "adv": 5.114, "adv_acc": 0.18, "gn": 2.23, "B": 13, "lr": 9.965e-06, "skips": 2, "oom": 0, "min": 1.09}
+{"step": 225, "ce": 3.6002, "acc": 0.2083, "adv": 4.485, "adv_acc": 0.172, "gn": 2.06, "B": 28, "lr": 9.956e-06, "skips": 2, "oom": 0, "min": 1.22}
+{"step": 250, "ce": 3.831, "acc": 0.175, "adv": 5.128, "adv_acc": 0.004, "gn": 2.1, "B": 13, "lr": 9.945e-06, "skips": 2, "oom": 0, "min": 1.35}
+{"step": 275, "ce": 3.5115, "acc": 0.2121, "adv": 5.416, "adv_acc": 0.0, "gn": 2.1, "B": 17, "lr": 9.933e-06, "skips": 2, "oom": 0, "min": 1.49}
+{"step": 300, "ce": 3.6931, "acc": 0.1905, "adv": 5.637, "adv_acc": 0.011, "gn": 1.87, "B": 14, "lr": 9.921e-06, "skips": 2, "oom": 0, "min": 1.61}
+{"step": 325, "ce": 3.933, "acc": 0.1594, "adv": 6.282, "adv_acc": 0.022, "gn": 2.38, "B": 41, "lr": 9.907e-06, "skips": 2, "oom": 0, "min": 1.73}
+{"step": 350, "ce": 3.3591, "acc": 0.2514, "adv": 4.969, "adv_acc": 0.001, "gn": 2.06, "B": 10, "lr": 9.892e-06, "skips": 2, "oom": 0, "min": 1.85}
+{"step": 375, "ce": 3.6679, "acc": 0.1897, "adv": 5.176, "adv_acc": 0.12, "gn": 1.9, "B": 15, "lr": 9.876e-06, "skips": 2, "oom": 0, "min": 1.97}
+{"step": 400, "ce": 3.8537, "acc": 0.1769, "adv": 5.109, "adv_acc": 0.024, "gn": 1.81, "B": 23, "lr": 9.859e-06, "skips": 2, "oom": 0, "min": 2.07}
+{"step": 425, "ce": 3.6942, "acc": 0.1925, "adv": 4.691, "adv_acc": 0.125, "gn": 2.17, "B": 26, "lr": 9.841e-06, "skips": 2, "oom": 0, "min": 2.2}
+{"step": 450, "ce": 3.9306, "acc": 0.1615, "adv": 6.244, "adv_acc": 0.022, "gn": 2.23, "B": 44, "lr": 9.822e-06, "skips": 2, "oom": 0, "min": 2.33}
+{"step": 475, "ce": 3.343, "acc": 0.2304, "adv": 4.348, "adv_acc": 0.115, "gn": 2.17, "B": 16, "lr": 9.802e-06, "skips": 2, "oom": 0, "min": 2.45}
+{"step": 500, "ce": 3.6741, "acc": 0.1936, "adv": 4.489, "adv_acc": 0.025, "gn": 1.83, "B": 21, "lr": 9.781e-06, "skips": 3, "oom": 0, "min": 2.57}
+{"step": 500, "val": {"np": 3.507689257786406, "p": 3.512645748616818, "anchor_p": 3.1165452003479004, "mean": 3.510167503201612}, "best": true}
+{"step": 525, "ce": 3.7693, "acc": 0.1915, "adv": 8.908, "adv_acc": 0.0, "gn": 2.37, "B": 59, "lr": 9.758e-06, "skips": 3, "oom": 0, "min": 2.77}
+{"step": 550, "ce": 3.5181, "acc": 0.2223, "adv": 4.695, "adv_acc": 0.103, "gn": 2.16, "B": 14, "lr": 9.735e-06, "skips": 4, "oom": 0, "min": 2.89}
+{"step": 575, "ce": 3.5077, "acc": 0.2158, "adv": 4.792, "adv_acc": 0.0, "gn": 3.13, "B": 7, "lr": 9.71e-06, "skips": 4, "oom": 0, "min": 3.01}
+{"step": 600, "ce": 3.4617, "acc": 0.2337, "adv": 4.699, "adv_acc": 0.032, "gn": 2.14, "B": 10, "lr": 9.685e-06, "skips": 4, "oom": 0, "min": 3.14}
+{"step": 625, "ce": 3.5601, "acc": 0.1991, "adv": 4.77, "adv_acc": 0.021, "gn": 2.13, "B": 24, "lr": 9.659e-06, "skips": 5, "oom": 0, "min": 3.25}
+{"step": 650, "ce": 3.3011, "acc": 0.2631, "adv": 4.122, "adv_acc": 0.011, "gn": 1.85, "B": 14, "lr": 9.631e-06, "skips": 5, "oom": 0, "min": 3.38}
+{"step": 675, "ce": 3.5935, "acc": 0.2031, "adv": 4.653, "adv_acc": 0.033, "gn": 1.88, "B": 19, "lr": 9.602e-06, "skips": 5, "oom": 0, "min": 3.5}
+{"step": 700, "ce": 3.2415, "acc": 0.2494, "adv": 4.693, "adv_acc": 0.003, "gn": 2.14, "B": 7, "lr": 9.573e-06, "skips": 5, "oom": 0, "min": 3.63}
+{"step": 725, "ce": 3.2985, "acc": 0.2415, "adv": 4.372, "adv_acc": 0.129, "gn": 1.99, "B": 11, "lr": 9.542e-06, "skips": 5, "oom": 0, "min": 3.76}
+{"step": 750, "ce": 3.5693, "acc": 0.1964, "adv": 4.211, "adv_acc": 0.047, "gn": 2.05, "B": 27, "lr": 9.511e-06, "skips": 5, "oom": 0, "min": 3.88}
+{"step": 775, "ce": 3.26, "acc": 0.2431, "adv": 4.574, "adv_acc": 0.007, "gn": 2.29, "B": 8, "lr": 9.478e-06, "skips": 5, "oom": 0, "min": 4.01}
+{"step": 800, "ce": 3.3265, "acc": 0.2363, "adv": 4.495, "adv_acc": 0.047, "gn": 2.32, "B": 9, "lr": 9.445e-06, "skips": 5, "oom": 0, "min": 4.13}
+{"step": 825, "ce": 3.7204, "acc": 0.1812, "adv": 7.24, "adv_acc": 0.002, "gn": 2.3, "B": 29, "lr": 9.41e-06, "skips": 5, "oom": 0, "min": 4.27}
+{"step": 850, "ce": 3.5532, "acc": 0.2079, "adv": 15.118, "adv_acc": 0.006, "gn": 3.38, "B": 11, "lr": 9.375e-06, "skips": 5, "oom": 0, "min": 4.38}
+{"step": 875, "ce": 3.9259, "acc": 0.1713, "adv": 40.286, "adv_acc": 0.017, "gn": 4.35, "B": 41, "lr": 9.338e-06, "skips": 5, "oom": 0, "min": 4.51}
+{"step": 900, "ce": 3.9721, "acc": 0.1658, "adv": 43.79, "adv_acc": 0.02, "gn": 6.45, "B": 51, "lr": 9.301e-06, "skips": 5, "oom": 0, "min": 4.62}
+{"step": 925, "ce": 3.4603, "acc": 0.2349, "adv": 10.369, "adv_acc": 0.0, "gn": 2.68, "B": 14, "lr": 9.263e-06, "skips": 5, "oom": 0, "min": 4.74}
+{"step": 950, "ce": 3.5721, "acc": 0.2092, "adv": 9.951, "adv_acc": 0.032, "gn": 2.45, "B": 20, "lr": 9.223e-06, "skips": 5, "oom": 0, "min": 4.85}
+{"step": 975, "ce": 3.7512, "acc": 0.1921, "adv": 38.525, "adv_acc": 0.0, "gn": 5.27, "B": 65, "lr": 9.183e-06, "skips": 5, "oom": 0, "min": 4.97}
+{"step": 1000, "ce": 3.4875, "acc": 0.2139, "adv": 9.408, "adv_acc": 0.104, "gn": 2.57, "B": 19, "lr": 9.142e-06, "skips": 5, "oom": 0, "min": 5.1}
+{"step": 1000, "val": {"np": 3.4977280627632763, "p": 3.5005001353906886, "anchor_p": 3.1229519844055176, "mean": 3.4991140990769827}, "best": true}
+{"step": 1025, "ce": 3.4877, "acc": 0.2195, "adv": 15.951, "adv_acc": 0.0, "gn": 3.29, "B": 16, "lr": 9.1e-06, "skips": 6, "oom": 0, "min": 5.31}
+{"step": 1050, "ce": 3.4113, "acc": 0.2332, "adv": 19.778, "adv_acc": 0.08, "gn": 3.95, "B": 9, "lr": 9.057e-06, "skips": 6, "oom": 0, "min": 5.43}
+{"step": 1075, "ce": 3.1604, "acc": 0.2612, "adv": 10.481, "adv_acc": 0.001, "gn": 3.24, "B": 7, "lr": 9.014e-06, "skips": 6, "oom": 0, "min": 5.54}
+{"step": 1100, "ce": 3.9686, "acc": 0.1677, "adv": 33.668, "adv_acc": 0.0, "gn": 4.65, "B": 58, "lr": 8.969e-06, "skips": 6, "oom": 0, "min": 5.66}
+{"step": 1125, "ce": 3.8179, "acc": 0.1783, "adv": 16.777, "adv_acc": 0.002, "gn": 3.08, "B": 36, "lr": 8.924e-06, "skips": 6, "oom": 0, "min": 5.78}
+{"step": 1150, "ce": 3.6925, "acc": 0.1956, "adv": 10.705, "adv_acc": 0.012, "gn": 2.96, "B": 14, "lr": 8.877e-06, "skips": 6, "oom": 0, "min": 5.89}
+{"step": 1175, "ce": 3.746, "acc": 0.1831, "adv": 11.042, "adv_acc": 0.001, "gn": 2.63, "B": 18, "lr": 8.83e-06, "skips": 6, "oom": 0, "min": 6.01}
+{"step": 1200, "ce": 3.6621, "acc": 0.195, "adv": 8.882, "adv_acc": 0.0, "gn": 2.38, "B": 10, "lr": 8.782e-06, "skips": 6, "oom": 0, "min": 6.14}
+{"step": 1225, "ce": 3.4611, "acc": 0.2234, "adv": 12.372, "adv_acc": 0.047, "gn": 2.56, "B": 18, "lr": 8.734e-06, "skips": 6, "oom": 0, "min": 6.26}
+{"step": 1250, "ce": 3.7915, "acc": 0.1761, "adv": 9.404, "adv_acc": 0.045, "gn": 2.54, "B": 13, "lr": 8.684e-06, "skips": 6, "oom": 0, "min": 6.37}
+{"step": 1275, "ce": 3.766, "acc": 0.1962, "adv": 22.222, "adv_acc": 0.0, "gn": 3.98, "B": 57, "lr": 8.634e-06, "skips": 6, "oom": 0, "min": 6.49}
+{"step": 1300, "ce": 3.772, "acc": 0.1855, "adv": 9.375, "adv_acc": 0.001, "gn": 2.33, "B": 22, "lr": 8.583e-06, "skips": 6, "oom": 0, "min": 6.61}
+{"step": 1325, "ce": 3.2919, "acc": 0.2332, "adv": 4.917, "adv_acc": 0.053, "gn": 2.88, "B": 7, "lr": 8.531e-06, "skips": 6, "oom": 0, "min": 6.73}
+{"step": 1350, "ce": 3.3606, "acc": 0.2269, "adv": 8.513, "adv_acc": 0.001, "gn": 2.57, "B": 13, "lr": 8.478e-06, "skips": 6, "oom": 0, "min": 6.85}
+{"step": 1375, "ce": 3.4161, "acc": 0.2352, "adv": 6.922, "adv_acc": 0.013, "gn": 2.26, "B": 14, "lr": 8.425e-06, "skips": 6, "oom": 0, "min": 6.96}
+{"step": 1400, "ce": 3.325, "acc": 0.2363, "adv": 8.211, "adv_acc": 0.007, "gn": 2.91, "B": 9, "lr": 8.371e-06, "skips": 6, "oom": 0, "min": 7.08}
+{"step": 1425, "ce": 3.7772, "acc": 0.1893, "adv": 22.674, "adv_acc": 0.0, "gn": 4.1, "B": 61, "lr": 8.316e-06, "skips": 6, "oom": 0, "min": 7.2}
+{"step": 1450, "ce": 3.5065, "acc": 0.2128, "adv": 7.335, "adv_acc": 0.032, "gn": 2.11, "B": 29, "lr": 8.26e-06, "skips": 6, "oom": 0, "min": 7.33}
+{"step": 1475, "ce": 3.5474, "acc": 0.2084, "adv": 6.431, "adv_acc": 0.009, "gn": 2.18, "B": 26, "lr": 8.204e-06, "skips": 6, "oom": 0, "min": 7.44}
+{"step": 1500, "ce": 3.7764, "acc": 0.1759, "adv": 11.695, "adv_acc": 0.041, "gn": 2.27, "B": 43, "lr": 8.147e-06, "skips": 7, "oom": 0, "min": 7.56}
+{"step": 1500, "val": {"np": 3.483489041219705, "p": 3.4874038346815577, "anchor_p": 3.090611457824707, "mean": 3.4854464379506314}, "best": true}
+{"step": 1525, "ce": 3.5611, "acc": 0.2128, "adv": 6.366, "adv_acc": 0.017, "gn": 2.16, "B": 12, "lr": 8.09e-06, "skips": 7, "oom": 0, "min": 7.76}
+{"step": 1550, "ce": 3.571, "acc": 0.2182, "adv": 11.321, "adv_acc": 0.004, "gn": 3.66, "B": 76, "lr": 8.032e-06, "skips": 7, "oom": 0, "min": 7.89}
+{"step": 1575, "ce": 3.4323, "acc": 0.2329, "adv": 6.34, "adv_acc": 0.015, "gn": 2.26, "B": 10, "lr": 7.973e-06, "skips": 7, "oom": 0, "min": 7.99}
+{"step": 1600, "ce": 3.7183, "acc": 0.1903, "adv": 11.932, "adv_acc": 0.005, "gn": 3.07, "B": 51, "lr": 7.914e-06, "skips": 7, "oom": 0, "min": 8.1}
+{"step": 1625, "ce": 3.4348, "acc": 0.2274, "adv": 6.191, "adv_acc": 0.005, "gn": 2.24, "B": 12, "lr": 7.854e-06, "skips": 7, "oom": 0, "min": 8.23}
+{"step": 1650, "ce": 3.5439, "acc": 0.1908, "adv": 7.033, "adv_acc": 0.0, "gn": 2.77, "B": 6, "lr": 7.793e-06, "skips": 7, "oom": 0, "min": 8.35}
+{"step": 1675, "ce": 3.7204, "acc": 0.1895, "adv": 6.672, "adv_acc": 0.041, "gn": 1.98, "B": 27, "lr": 7.732e-06, "skips": 7, "oom": 0, "min": 8.47}
+{"step": 1700, "ce": 3.6902, "acc": 0.1898, "adv": 6.503, "adv_acc": 0.031, "gn": 2.17, "B": 29, "lr": 7.67e-06, "skips": 8, "oom": 0, "min": 8.6}
+{"step": 1725, "ce": 3.5412, "acc": 0.2078, "adv": 6.138, "adv_acc": 0.022, "gn": 2.09, "B": 21, "lr": 7.608e-06, "skips": 8, "oom": 0, "min": 8.73}
+{"step": 1750, "ce": 3.5848, "acc": 0.208, "adv": 6.601, "adv_acc": 0.056, "gn": 2.2, "B": 10, "lr": 7.545e-06, "skips": 8, "oom": 0, "min": 8.84}
+{"step": 1775, "ce": 3.4334, "acc": 0.2558, "adv": 18.173, "adv_acc": 0.0, "gn": 3.74, "B": 69, "lr": 7.482e-06, "skips": 8, "oom": 0, "min": 8.95}
+{"step": 1800, "ce": 3.3952, "acc": 0.2212, "adv": 4.707, "adv_acc": 0.064, "gn": 2.3, "B": 9, "lr": 7.419e-06, "skips": 8, "oom": 0, "min": 9.07}
+{"step": 1825, "ce": 3.6546, "acc": 0.2024, "adv": 6.154, "adv_acc": 0.044, "gn": 2.32, "B": 33, "lr": 7.354e-06, "skips": 8, "oom": 0, "min": 9.19}
+{"step": 1850, "ce": 3.5126, "acc": 0.2026, "adv": 6.185, "adv_acc": 0.01, "gn": 2.16, "B": 26, "lr": 7.29e-06, "skips": 8, "oom": 0, "min": 9.3}
+{"step": 1875, "ce": 3.0871, "acc": 0.2778, "adv": 4.731, "adv_acc": 0.098, "gn": 2.17, "B": 10, "lr": 7.225e-06, "skips": 8, "oom": 0, "min": 9.41}
+{"step": 1900, "ce": 3.1692, "acc": 0.2457, "adv": 5.498, "adv_acc": 0.015, "gn": 2.32, "B": 6, "lr": 7.159e-06, "skips": 9, "oom": 0, "min": 9.53}
+{"step": 1925, "ce": 3.6654, "acc": 0.1816, "adv": 11.98, "adv_acc": 0.001, "gn": 2.36, "B": 41, "lr": 7.093e-06, "skips": 9, "oom": 0, "min": 9.65}
+{"step": 1950, "ce": 3.691, "acc": 0.1926, "adv": 15.862, "adv_acc": 0.0, "gn": 3.8, "B": 67, "lr": 7.027e-06, "skips": 9, "oom": 0, "min": 9.76}
+{"step": 1975, "ce": 3.569, "acc": 0.2165, "adv": 5.811, "adv_acc": 0.034, "gn": 2.0, "B": 16, "lr": 6.96e-06, "skips": 9, "oom": 0, "min": 9.89}
+{"step": 2000, "ce": 3.6569, "acc": 0.2025, "adv": 5.672, "adv_acc": 0.011, "gn": 2.62, "B": 24, "lr": 6.893e-06, "skips": 9, "oom": 0, "min": 10.01}
+{"step": 2000, "val": {"np": 3.4673069890236623, "p": 3.473822727265498, "anchor_p": 3.0952401161193848, "mean": 3.47056485814458}, "best": true}
+{"step": 2025, "ce": 3.5417, "acc": 0.2055, "adv": 6.049, "adv_acc": 0.031, "gn": 2.45, "B": 10, "lr": 6.826e-06, "skips": 9, "oom": 0, "min": 10.2}
+{"step": 2050, "ce": 3.3296, "acc": 0.2364, "adv": 5.457, "adv_acc": 0.018, "gn": 2.29, "B": 8, "lr": 6.758e-06, "skips": 9, "oom": 0, "min": 10.32}
+{"step": 2075, "ce": 3.39, "acc": 0.2286, "adv": 5.628, "adv_acc": 0.001, "gn": 2.08, "B": 13, "lr": 6.69e-06, "skips": 9, "oom": 0, "min": 10.43}
+{"step": 2100, "ce": 3.7423, "acc": 0.1898, "adv": 10.36, "adv_acc": 0.003, "gn": 2.33, "B": 43, "lr": 6.622e-06, "skips": 9, "oom": 0, "min": 10.56}
+{"step": 2125, "ce": 3.6225, "acc": 0.2003, "adv": 6.223, "adv_acc": 0.005, "gn": 1.99, "B": 13, "lr": 6.553e-06, "skips": 9, "oom": 0, "min": 10.68}
+{"step": 2150, "ce": 3.4632, "acc": 0.2213, "adv": 5.662, "adv_acc": 0.0, "gn": 1.88, "B": 16, "lr": 6.484e-06, "skips": 9, "oom": 0, "min": 10.8}
+{"step": 2175, "ce": 3.2873, "acc": 0.2562, "adv": 5.891, "adv_acc": 0.004, "gn": 2.14, "B": 10, "lr": 6.415e-06, "skips": 9, "oom": 0, "min": 10.92}
+{"step": 2200, "ce": 3.9016, "acc": 0.1671, "adv": 8.882, "adv_acc": 0.005, "gn": 2.36, "B": 41, "lr": 6.346e-06, "skips": 9, "oom": 0, "min": 11.03}
+{"step": 2225, "ce": 2.9077, "acc": 0.3194, "adv": 5.976, "adv_acc": 0.013, "gn": 2.83, "B": 7, "lr": 6.276e-06, "skips": 9, "oom": 0, "min": 11.15}
+{"step": 2250, "ce": 3.6433, "acc": 0.2093, "adv": 13.079, "adv_acc": 0.001, "gn": 3.92, "B": 54, "lr": 6.207e-06, "skips": 9, "oom": 0, "min": 11.26}
+{"step": 2275, "ce": 3.4893, "acc": 0.2168, "adv": 5.574, "adv_acc": 0.06, "gn": 1.87, "B": 14, "lr": 6.137e-06, "skips": 10, "oom": 0, "min": 11.38}
+{"step": 2300, "ce": 3.704, "acc": 0.1807, "adv": 10.287, "adv_acc": 0.0, "gn": 2.77, "B": 50, "lr": 6.067e-06, "skips": 10, "oom": 0, "min": 11.5}
+{"step": 2325, "ce": 3.8156, "acc": 0.1893, "adv": 13.466, "adv_acc": 0.0, "gn": 3.07, "B": 55, "lr": 5.997e-06, "skips": 10, "oom": 0, "min": 11.61}
+{"step": 2350, "ce": 3.5805, "acc": 0.2117, "adv": 14.332, "adv_acc": 0.005, "gn": 3.27, "B": 68, "lr": 5.926e-06, "skips": 10, "oom": 0, "min": 11.74}
+{"step": 2375, "ce": 3.3738, "acc": 0.2044, "adv": 6.482, "adv_acc": 0.0, "gn": 2.93, "B": 6, "lr": 5.856e-06, "skips": 10, "oom": 0, "min": 11.87}
+{"step": 2400, "ce": 3.6756, "acc": 0.1828, "adv": 6.471, "adv_acc": 0.005, "gn": 1.93, "B": 12, "lr": 5.785e-06, "skips": 10, "oom": 0, "min": 11.99}
+{"step": 2425, "ce": 3.3128, "acc": 0.2365, "adv": 6.476, "adv_acc": 0.013, "gn": 1.89, "B": 10, "lr": 5.715e-06, "skips": 10, "oom": 0, "min": 12.1}
+{"step": 2450, "ce": 3.4673, "acc": 0.2181, "adv": 6.946, "adv_acc": 0.015, "gn": 2.02, "B": 33, "lr": 5.644e-06, "skips": 10, "oom": 0, "min": 12.22}
+{"step": 2475, "ce": 3.5298, "acc": 0.2094, "adv": 6.09, "adv_acc": 0.039, "gn": 1.71, "B": 19, "lr": 5.574e-06, "skips": 10, "oom": 0, "min": 12.32}
+{"step": 2500, "ce": 3.2715, "acc": 0.2515, "adv": 6.632, "adv_acc": 0.121, "gn": 1.87, "B": 17, "lr": 5.503e-06, "skips": 10, "oom": 0, "min": 12.43}
+{"step": 2500, "val": {"np": 3.455619011329129, "p": 3.4601697774197455, "anchor_p": 3.0822691917419434, "mean": 3.4578943943744376}, "best": true}
+{"step": 2525, "ce": 3.7506, "acc": 0.183, "adv": 15.727, "adv_acc": 0.002, "gn": 2.97, "B": 55, "lr": 5.432e-06, "skips": 10, "oom": 0, "min": 12.63}
+{"step": 2550, "ce": 3.2503, "acc": 0.2422, "adv": 6.394, "adv_acc": 0.012, "gn": 2.1, "B": 10, "lr": 5.361e-06, "skips": 11, "oom": 0, "min": 12.75}
+{"step": 2575, "ce": 3.3457, "acc": 0.2401, "adv": 6.05, "adv_acc": 0.119, "gn": 1.93, "B": 17, "lr": 5.291e-06, "skips": 11, "oom": 0, "min": 12.86}
+{"step": 2600, "ce": 3.6437, "acc": 0.1846, "adv": 6.671, "adv_acc": 0.061, "gn": 1.87, "B": 29, "lr": 5.22e-06, "skips": 11, "oom": 0, "min": 12.97}
+{"step": 2625, "ce": 3.5463, "acc": 0.2367, "adv": 14.848, "adv_acc": 0.003, "gn": 3.72, "B": 74, "lr": 5.15e-06, "skips": 11, "oom": 0, "min": 13.09}
+{"step": 2650, "ce": 3.3847, "acc": 0.2314, "adv": 6.816, "adv_acc": 0.016, "gn": 2.39, "B": 9, "lr": 5.079e-06, "skips": 11, "oom": 0, "min": 13.2}
+{"step": 2675, "ce": 3.5453, "acc": 0.2041, "adv": 6.815, "adv_acc": 0.003, "gn": 2.07, "B": 23, "lr": 5.009e-06, "skips": 11, "oom": 0, "min": 13.31}
+{"step": 2700, "ce": 3.3028, "acc": 0.2437, "adv": 7.003, "adv_acc": 0.128, "gn": 1.87, "B": 15, "lr": 4.939e-06, "skips": 11, "oom": 0, "min": 13.43}
+{"step": 2725, "ce": 3.8053, "acc": 0.1819, "adv": 9.202, "adv_acc": 0.03, "gn": 2.73, "B": 42, "lr": 4.869e-06, "skips": 11, "oom": 0, "min": 13.54}
+{"step": 2750, "ce": 3.586, "acc": 0.2259, "adv": 10.705, "adv_acc": 0.038, "gn": 3.32, "B": 76, "lr": 4.799e-06, "skips": 12, "oom": 0, "min": 13.66}
+{"step": 2775, "ce": 3.242, "acc": 0.2543, "adv": 7.569, "adv_acc": 0.147, "gn": 1.83, "B": 14, "lr": 4.729e-06, "skips": 12, "oom": 0, "min": 13.77}
+{"step": 2800, "ce": 3.5247, "acc": 0.2002, "adv": 7.495, "adv_acc": 0.0, "gn": 1.98, "B": 12, "lr": 4.66e-06, "skips": 12, "oom": 0, "min": 13.88}
+{"step": 2825, "ce": 3.5414, "acc": 0.2004, "adv": 7.73, "adv_acc": 0.06, "gn": 2.6, "B": 40, "lr": 4.59e-06, "skips": 12, "oom": 0, "min": 13.99}
+{"step": 2850, "ce": 3.2731, "acc": 0.2393, "adv": 7.595, "adv_acc": 0.051, "gn": 1.9, "B": 21, "lr": 4.521e-06, "skips": 12, "oom": 0, "min": 14.1}
+{"step": 2875, "ce": 3.427, "acc": 0.2152, "adv": 7.72, "adv_acc": 0.0, "gn": 1.96, "B": 10, "lr": 4.452e-06, "skips": 13, "oom": 0, "min": 14.23}
+{"step": 2900, "ce": 3.4472, "acc": 0.2124, "adv": 7.726, "adv_acc": 0.0, "gn": 1.9, "B": 14, "lr": 4.384e-06, "skips": 13, "oom": 0, "min": 14.34}
+{"step": 2925, "ce": 3.3561, "acc": 0.228, "adv": 7.77, "adv_acc": 0.139, "gn": 1.95, "B": 17, "lr": 4.315e-06, "skips": 13, "oom": 0, "min": 14.45}
+{"step": 2950, "ce": 3.6581, "acc": 0.2103, "adv": 8.242, "adv_acc": 0.001, "gn": 2.15, "B": 53, "lr": 4.247e-06, "skips": 13, "oom": 0, "min": 14.58}
+{"step": 2975, "ce": 3.5687, "acc": 0.2064, "adv": 7.748, "adv_acc": 0.0, "gn": 2.09, "B": 11, "lr": 4.18e-06, "skips": 13, "oom": 0, "min": 14.69}
+{"step": 3000, "ce": 3.8188, "acc": 0.1662, "adv": 8.003, "adv_acc": 0.01, "gn": 11.11, "B": 52, "lr": 4.112e-06, "skips": 13, "oom": 0, "min": 14.81}
+{"step": 3000, "val": {"np": 3.458494051271619, "p": 3.4631192987050605, "anchor_p": 3.0834555625915527, "mean": 3.4608066749883397}, "best": false}
+{"step": 3025, "ce": 3.8608, "acc": 0.1688, "adv": 7.726, "adv_acc": 0.001, "gn": 2.45, "B": 47, "lr": 4.045e-06, "skips": 14, "oom": 0, "min": 14.96}
+{"step": 3050, "ce": 3.5531, "acc": 0.2074, "adv": 7.65, "adv_acc": 0.193, "gn": 2.03, "B": 13, "lr": 3.978e-06, "skips": 14, "oom": 0, "min": 15.07}
+{"step": 3075, "ce": 3.6485, "acc": 0.2358, "adv": 8.172, "adv_acc": 0.0, "gn": 2.18, "B": 54, "lr": 3.912e-06, "skips": 14, "oom": 0, "min": 15.19}
+{"step": 3100, "ce": 3.6049, "acc": 0.1975, "adv": 7.879, "adv_acc": 0.015, "gn": 1.97, "B": 43, "lr": 3.846e-06, "skips": 14, "oom": 0, "min": 15.3}
+{"step": 3125, "ce": 3.4728, "acc": 0.2281, "adv": 7.707, "adv_acc": 0.002, "gn": 1.97, "B": 15, "lr": 3.781e-06, "skips": 14, "oom": 0, "min": 15.4}
+{"step": 3150, "ce": 3.1627, "acc": 0.248, "adv": 7.661, "adv_acc": 0.001, "gn": 2.15, "B": 7, "lr": 3.715e-06, "skips": 14, "oom": 0, "min": 15.51}
+{"step": 3175, "ce": 3.632, "acc": 0.1925, "adv": 8.086, "adv_acc": 0.0, "gn": 2.32, "B": 61, "lr": 3.651e-06, "skips": 15, "oom": 0, "min": 15.64}
+{"step": 3200, "ce": 3.1466, "acc": 0.2697, "adv": 7.568, "adv_acc": 0.001, "gn": 2.21, "B": 8, "lr": 3.587e-06, "skips": 15, "oom": 0, "min": 15.76}
+{"step": 3225, "ce": 3.4873, "acc": 0.2071, "adv": 7.532, "adv_acc": 0.0, "gn": 2.03, "B": 15, "lr": 3.523e-06, "skips": 15, "oom": 0, "min": 15.87}
+{"step": 3250, "ce": 3.4786, "acc": 0.2021, "adv": 7.501, "adv_acc": 0.078, "gn": 1.92, "B": 24, "lr": 3.46e-06, "skips": 15, "oom": 0, "min": 15.97}
+{"step": 3275, "ce": 3.8398, "acc": 0.1564, "adv": 7.779, "adv_acc": 0.0, "gn": 2.0, "B": 45, "lr": 3.397e-06, "skips": 15, "oom": 0, "min": 16.09}
+{"step": 3300, "ce": 3.4721, "acc": 0.1948, "adv": 7.401, "adv_acc": 0.0, "gn": 2.6, "B": 6, "lr": 3.335e-06, "skips": 15, "oom": 0, "min": 16.2}
+{"step": 3325, "ce": 3.4505, "acc": 0.2151, "adv": 7.363, "adv_acc": 0.037, "gn": 2.08, "B": 19, "lr": 3.273e-06, "skips": 17, "oom": 0, "min": 16.32}
+{"step": 3350, "ce": 3.6933, "acc": 0.1805, "adv": 8.033, "adv_acc": 0.0, "gn": 2.17, "B": 54, "lr": 3.212e-06, "skips": 17, "oom": 0, "min": 16.43}
+{"step": 3375, "ce": 3.6515, "acc": 0.1935, "adv": 7.364, "adv_acc": 0.001, "gn": 1.94, "B": 26, "lr": 3.151e-06, "skips": 17, "oom": 0, "min": 16.54}
+{"step": 3400, "ce": 3.3099, "acc": 0.2384, "adv": 7.2, "adv_acc": 0.241, "gn": 1.99, "B": 18, "lr": 3.091e-06, "skips": 18, "oom": 0, "min": 16.66}
+{"step": 3425, "ce": 3.4313, "acc": 0.2122, "adv": 7.262, "adv_acc": 0.0, "gn": 2.29, "B": 11, "lr": 3.032e-06, "skips": 18, "oom": 0, "min": 16.77}
+{"step": 3450, "ce": 3.1878, "acc": 0.2464, "adv": 7.173, "adv_acc": 0.109, "gn": 1.89, "B": 18, "lr": 2.973e-06, "skips": 18, "oom": 0, "min": 16.88}
+{"step": 3475, "ce": 3.1116, "acc": 0.2702, "adv": 7.052, "adv_acc": 0.115, "gn": 2.0, "B": 12, "lr": 2.915e-06, "skips": 18, "oom": 0, "min": 16.98}
+{"step": 3500, "ce": 3.69, "acc": 0.183, "adv": 7.387, "adv_acc": 0.001, "gn": 2.04, "B": 44, "lr": 2.857e-06, "skips": 18, "oom": 0, "min": 17.1}
+{"step": 3500, "val": {"np": 3.450007374201225, "p": 3.453779433371578, "anchor_p": 3.084826946258545, "mean": 3.4518934037864017}, "best": true}
+{"step": 3525, "ce": 3.8386, "acc": 0.1708, "adv": 7.872, "adv_acc": 0.0, "gn": 2.22, "B": 59, "lr": 2.8e-06, "skips": 18, "oom": 0, "min": 17.29}
+{"step": 3550, "ce": 3.3008, "acc": 0.2418, "adv": 6.87, "adv_acc": 0.256, "gn": 2.36, "B": 18, "lr": 2.744e-06, "skips": 18, "oom": 0, "min": 17.39}
+{"step": 3575, "ce": 3.2114, "acc": 0.2499, "adv": 6.952, "adv_acc": 0.143, "gn": 2.01, "B": 25, "lr": 2.689e-06, "skips": 18, "oom": 0, "min": 17.49}
+{"step": 3600, "ce": 3.631, "acc": 0.2007, "adv": 7.465, "adv_acc": 0.0, "gn": 2.28, "B": 48, "lr": 2.634e-06, "skips": 18, "oom": 0, "min": 17.59}
+{"step": 3625, "ce": 3.0395, "acc": 0.2746, "adv": 6.853, "adv_acc": 0.365, "gn": 2.04, "B": 14, "lr": 2.58e-06, "skips": 19, "oom": 0, "min": 17.71}
+{"step": 3650, "ce": 3.2907, "acc": 0.2407, "adv": 6.957, "adv_acc": 0.059, "gn": 2.04, "B": 19, "lr": 2.526e-06, "skips": 19, "oom": 0, "min": 17.82}
+{"step": 3675, "ce": 3.4421, "acc": 0.219, "adv": 6.903, "adv_acc": 0.078, "gn": 2.07, "B": 18, "lr": 2.474e-06, "skips": 20, "oom": 0, "min": 17.94}
+{"step": 3700, "ce": 3.3729, "acc": 0.2236, "adv": 6.87, "adv_acc": 0.087, "gn": 2.08, "B": 11, "lr": 2.422e-06, "skips": 21, "oom": 0, "min": 18.05}
+{"step": 3725, "ce": 3.8656, "acc": 0.1712, "adv": 7.638, "adv_acc": 0.001, "gn": 2.12, "B": 48, "lr": 2.37e-06, "skips": 21, "oom": 0, "min": 18.16}
+{"step": 3750, "ce": 3.1599, "acc": 0.251, "adv": 6.81, "adv_acc": 0.002, "gn": 2.31, "B": 9, "lr": 2.32e-06, "skips": 21, "oom": 0, "min": 18.26}
+{"step": 3775, "ce": 3.597, "acc": 0.2018, "adv": 7.29, "adv_acc": 0.0, "gn": 2.19, "B": 45, "lr": 2.27e-06, "skips": 21, "oom": 0, "min": 18.35}
+{"step": 3800, "ce": 3.4677, "acc": 0.2096, "adv": 7.013, "adv_acc": 0.021, "gn": 2.12, "B": 38, "lr": 2.222e-06, "skips": 21, "oom": 0, "min": 18.46}
+{"step": 3825, "ce": 3.8072, "acc": 0.162, "adv": 7.709, "adv_acc": 0.0, "gn": 2.54, "B": 56, "lr": 2.174e-06, "skips": 21, "oom": 0, "min": 18.57}
+{"step": 3850, "ce": 3.3768, "acc": 0.2393, "adv": 6.691, "adv_acc": 0.07, "gn": 2.16, "B": 15, "lr": 2.126e-06, "skips": 22, "oom": 0, "min": 18.68}
+{"step": 3875, "ce": 3.4887, "acc": 0.2255, "adv": 8.748, "adv_acc": 0.001, "gn": 2.49, "B": 57, "lr": 2.08e-06, "skips": 22, "oom": 0, "min": 18.78}
+{"step": 3900, "ce": 3.3054, "acc": 0.2282, "adv": 6.594, "adv_acc": 0.169, "gn": 1.96, "B": 19, "lr": 2.034e-06, "skips": 23, "oom": 0, "min": 18.89}
+{"step": 3925, "ce": 3.366, "acc": 0.2343, "adv": 6.587, "adv_acc": 0.001, "gn": 1.94, "B": 25, "lr": 1.99e-06, "skips": 24, "oom": 0, "min": 19.01}
+{"step": 3950, "ce": 3.3814, "acc": 0.2161, "adv": 6.627, "adv_acc": 0.0, "gn": 2.39, "B": 8, "lr": 1.946e-06, "skips": 24, "oom": 0, "min": 19.13}
+{"step": 3975, "ce": 3.8303, "acc": 0.1731, "adv": 7.142, "adv_acc": 0.001, "gn": 2.05, "B": 40, "lr": 1.903e-06, "skips": 24, "oom": 0, "min": 19.25}
+{"step": 4000, "ce": 3.3377, "acc": 0.2298, "adv": 6.495, "adv_acc": 0.094, "gn": 2.11, "B": 14, "lr": 1.861e-06, "skips": 24, "oom": 0, "min": 19.34}
+{"step": 4000, "val": {"np": 3.4431558952269414, "p": 3.446689893057758, "anchor_p": 3.0781209468841553, "mean": 3.4449228941423495}, "best": true}
+{"step": 4025, "ce": 3.6726, "acc": 0.2001, "adv": 7.533, "adv_acc": 0.0, "gn": 2.45, "B": 48, "lr": 1.82e-06, "skips": 25, "oom": 0, "min": 19.52}
+{"step": 4050, "ce": 3.3085, "acc": 0.241, "adv": 6.482, "adv_acc": 0.087, "gn": 1.9, "B": 19, "lr": 1.78e-06, "skips": 26, "oom": 0, "min": 19.63}
+{"step": 4075, "ce": 3.3112, "acc": 0.2284, "adv": 6.608, "adv_acc": 0.0, "gn": 2.34, "B": 7, "lr": 1.74e-06, "skips": 26, "oom": 0, "min": 19.73}
+{"step": 4100, "ce": 3.2196, "acc": 0.2551, "adv": 6.462, "adv_acc": 0.062, "gn": 2.0, "B": 18, "lr": 1.702e-06, "skips": 27, "oom": 0, "min": 19.83}
+{"step": 4125, "ce": 3.3942, "acc": 0.2193, "adv": 6.398, "adv_acc": 0.072, "gn": 2.13, "B": 24, "lr": 1.665e-06, "skips": 27, "oom": 0, "min": 19.93}
+{"step": 4150, "ce": 3.5767, "acc": 0.1964, "adv": 6.64, "adv_acc": 0.015, "gn": 2.21, "B": 37, "lr": 1.628e-06, "skips": 27, "oom": 0, "min": 20.05}
+{"step": 4175, "ce": 3.1631, "acc": 0.2804, "adv": 8.278, "adv_acc": 0.0, "gn": 2.84, "B": 70, "lr": 1.593e-06, "skips": 27, "oom": 0, "min": 20.15}
+{"step": 4200, "ce": 3.5769, "acc": 0.2021, "adv": 6.396, "adv_acc": 0.045, "gn": 3.32, "B": 29, "lr": 1.558e-06, "skips": 27, "oom": 0, "min": 20.25}
+{"step": 4225, "ce": 3.4182, "acc": 0.228, "adv": 6.34, "adv_acc": 0.026, "gn": 1.91, "B": 23, "lr": 1.524e-06, "skips": 27, "oom": 0, "min": 20.34}
+{"step": 4250, "ce": 3.4323, "acc": 0.1862, "adv": 6.416, "adv_acc": 0.0, "gn": 4.24, "B": 4, "lr": 1.492e-06, "skips": 27, "oom": 0, "min": 20.46}
+{"step": 4275, "ce": 3.2959, "acc": 0.2454, "adv": 6.336, "adv_acc": 0.074, "gn": 2.14, "B": 10, "lr": 1.46e-06, "skips": 27, "oom": 0, "min": 20.56}
+{"step": 4300, "ce": 3.3964, "acc": 0.2241, "adv": 6.433, "adv_acc": 0.0, "gn": 2.44, "B": 9, "lr": 1.429e-06, "skips": 27, "oom": 0, "min": 20.67}
+{"step": 4325, "ce": 3.1359, "acc": 0.2704, "adv": 6.4, "adv_acc": 0.0, "gn": 2.44, "B": 8, "lr": 1.4e-06, "skips": 27, "oom": 0, "min": 20.78}
+{"step": 4350, "ce": 3.3125, "acc": 0.2263, "adv": 6.296, "adv_acc": 0.024, "gn": 2.0, "B": 21, "lr": 1.371e-06, "skips": 27, "oom": 0, "min": 20.87}
+{"step": 4375, "ce": 3.7765, "acc": 0.1722, "adv": 6.958, "adv_acc": 0.003, "gn": 2.27, "B": 44, "lr": 1.344e-06, "skips": 28, "oom": 0, "min": 20.98}
+{"step": 4400, "ce": 3.802, "acc": 0.1807, "adv": 6.978, "adv_acc": 0.048, "gn": 2.29, "B": 44, "lr": 1.317e-06, "skips": 28, "oom": 0, "min": 21.1}
+{"step": 4425, "ce": 3.9622, "acc": 0.1576, "adv": 7.175, "adv_acc": 0.001, "gn": 2.23, "B": 53, "lr": 1.292e-06, "skips": 28, "oom": 0, "min": 21.22}
+{"step": 4450, "ce": 3.5589, "acc": 0.1894, "adv": 6.487, "adv_acc": 0.121, "gn": 2.24, "B": 38, "lr": 1.267e-06, "skips": 28, "oom": 0, "min": 21.33}
+{"step": 4475, "ce": 3.5869, "acc": 0.1958, "adv": 6.216, "adv_acc": 0.091, "gn": 2.13, "B": 21, "lr": 1.244e-06, "skips": 28, "oom": 0, "min": 21.44}
+{"step": 4500, "ce": 3.5748, "acc": 0.2104, "adv": 8.296, "adv_acc": 0.001, "gn": 2.57, "B": 62, "lr": 1.221e-06, "skips": 28, "oom": 0, "min": 21.54}
+{"step": 4500, "val": {"np": 3.4411249983970817, "p": 3.4445827232509947, "anchor_p": 3.079256296157837, "mean": 3.442853860824038}, "best": true}
+{"step": 4525, "ce": 3.6671, "acc": 0.1935, "adv": 7.579, "adv_acc": 0.0, "gn": 2.41, "B": 61, "lr": 1.2e-06, "skips": 28, "oom": 0, "min": 21.73}
+{"step": 4550, "ce": 3.5396, "acc": 0.2035, "adv": 6.112, "adv_acc": 0.033, "gn": 2.28, "B": 28, "lr": 1.179e-06, "skips": 28, "oom": 0, "min": 21.84}
+{"step": 4575, "ce": 3.0832, "acc": 0.2803, "adv": 6.234, "adv_acc": 0.0, "gn": 2.13, "B": 8, "lr": 1.16e-06, "skips": 29, "oom": 0, "min": 21.94}
+{"step": 4600, "ce": 3.5924, "acc": 0.1956, "adv": 6.252, "adv_acc": 0.001, "gn": 2.17, "B": 29, "lr": 1.142e-06, "skips": 29, "oom": 0, "min": 22.05}
+{"step": 4625, "ce": 3.4295, "acc": 0.2224, "adv": 6.193, "adv_acc": 0.043, "gn": 2.09, "B": 20, "lr": 1.125e-06, "skips": 30, "oom": 0, "min": 22.17}
+{"step": 4650, "ce": 3.3532, "acc": 0.2302, "adv": 6.142, "adv_acc": 0.181, "gn": 2.09, "B": 22, "lr": 1.109e-06, "skips": 30, "oom": 0, "min": 22.29}
+{"step": 4675, "ce": 3.6147, "acc": 0.2032, "adv": 6.351, "adv_acc": 0.031, "gn": 2.01, "B": 35, "lr": 1.094e-06, "skips": 30, "oom": 0, "min": 22.39}
+{"step": 4700, "ce": 3.5387, "acc": 0.2016, "adv": 6.802, "adv_acc": 0.0, "gn": 2.13, "B": 43, "lr": 1.08e-06, "skips": 30, "oom": 0, "min": 22.51}
+{"step": 4725, "ce": 3.5325, "acc": 0.2115, "adv": 6.492, "adv_acc": 0.001, "gn": 2.37, "B": 41, "lr": 1.067e-06, "skips": 30, "oom": 0, "min": 22.62}
+{"step": 4750, "ce": 3.3788, "acc": 0.2414, "adv": 6.013, "adv_acc": 0.256, "gn": 1.93, "B": 16, "lr": 1.056e-06, "skips": 30, "oom": 0, "min": 22.73}
+{"step": 4775, "ce": 3.4268, "acc": 0.228, "adv": 6.181, "adv_acc": 0.0, "gn": 2.18, "B": 13, "lr": 1.045e-06, "skips": 30, "oom": 0, "min": 22.83}
+{"step": 4800, "ce": 3.4523, "acc": 0.2268, "adv": 6.0, "adv_acc": 0.077, "gn": 2.02, "B": 17, "lr": 1.036e-06, "skips": 30, "oom": 0, "min": 22.94}
+{"step": 4825, "ce": 3.5242, "acc": 0.2296, "adv": 7.628, "adv_acc": 0.044, "gn": 2.09, "B": 51, "lr": 1.027e-06, "skips": 30, "oom": 0, "min": 23.05}
+{"step": 4850, "ce": 3.1302, "acc": 0.2662, "adv": 5.944, "adv_acc": 0.324, "gn": 2.04, "B": 12, "lr": 1.02e-06, "skips": 30, "oom": 0, "min": 23.14}
+{"step": 4875, "ce": 3.4428, "acc": 0.2185, "adv": 6.03, "adv_acc": 0.165, "gn": 2.1, "B": 15, "lr": 1.014e-06, "skips": 30, "oom": 0, "min": 23.25}
+{"step": 4900, "ce": 3.1318, "acc": 0.2708, "adv": 6.019, "adv_acc": 0.0, "gn": 2.6, "B": 6, "lr": 1.009e-06, "skips": 30, "oom": 0, "min": 23.35}
+{"step": 4925, "ce": 3.5066, "acc": 0.2196, "adv": 6.043, "adv_acc": 0.039, "gn": 1.91, "B": 24, "lr": 1.005e-06, "skips": 30, "oom": 0, "min": 23.44}
+{"step": 4950, "ce": 3.4736, "acc": 0.2202, "adv": 7.73, "adv_acc": 0.0, "gn": 3.54, "B": 69, "lr": 1.002e-06, "skips": 30, "oom": 0, "min": 23.54}
+{"step": 4975, "ce": 3.0291, "acc": 0.2834, "adv": 5.771, "adv_acc": 0.001, "gn": 2.29, "B": 9, "lr": 1.001e-06, "skips": 30, "oom": 0, "min": 23.64}
+{"step": 5000, "ce": 3.4741, "acc": 0.2143, "adv": 5.99, "adv_acc": 0.0, "gn": 1.92, "B": 15, "lr": 1e-06, "skips": 30, "oom": 0, "min": 23.74}
+{"step": 5000, "val": {"np": 3.4417521021653465, "p": 3.445160358658831, "anchor_p": 3.078298807144165, "mean": 3.443456230412089}, "best": false}

infer.py

@@ -1,35 +1,58 @@
 #!/usr/bin/env python
-"""Dynamic Accent (Phase 2 preview): Nigerian multilingual, code-switching TTS on CosyVoice2-0.5B.
-
-Same inference path as the Phase 2 final samples: LM `p2_lm_B` (pause profile B) + anchor-SFT flow,
-anchor prompt clip 0 at the LM and the flow, single sample, sentence-level stitching, edge trim and
-internal-silence cap (0.6 s), 0.6 s between sentences.
-
-  python infer.py --text "<eng>this is dynamic accent,</eng> <pcm>e dey sound nice abi?</pcm>" --seed 0 --out out.wav
-
-Prosody v2 (experimental, `--pause_mode clause`; default is `model`): the text is cut at , ; : . ? ! and at language-switch points; each chunk
-is generated with the same LM/flow/prompt, its edge silence is trimmed (energy VAD, 30 ms kept), its speech is sped
-up pitch-preserving by `--speed` (WSOLA, default 1.0), and chunks are joined with set gaps from the real anchor's pause statistics:
-comma/;/: 0.40 s (+-0.05 jitter), sentence end 0.60 s, unpunctuated language switch 0.12 s, with 10 ms fades.
-`--pause_mode model --speed 1.0` reproduces the Phase 2 final samples (model-generated pauses, sentence stitching).
-
-Text: wrap each language span in <eng>, <pcm>, <yor>, <ibo> or <hau> tags. Untagged text is treated as English.
+"""Dynamic Accent: Nigerian multilingual, code-switching TTS on CosyVoice2-0.5B, with selectable model versions.
+
+Versions (`--version`):
+  phase2  LM `p2_lm_B` + anchor-SFT flow `p2_flow_sft`. Phase 2 sampler, sentence stitching, edge trim, internal
+          silences capped at 0.6 s, 0.6 s between sentences (exactly the Phase 2 final samples).
+  phase3  LM `p3_lm_full` (re-cut long-form data with natural pauses, ~102 h of Igbo) + the shared Phase 2 flow (see
+          settings.json and checkpoints/phase3/p3_lm_full/config.json). Silence-aware sampler (silence tokens exempt from the repetition check up
+          to a 1 s run), the whole paragraph in ONE pass (up to ~60 words per pass), no audio post-processing: every
+          pause and the pace come from the model.
+
+Voices (`--voice`): `anchor_male` (default; WAXAL Yoruba speaker 4), `waxal_ibo_5_female`, `waxal_yor_5_female`
+(see --list_voices; recommended: phase3 with a female voice, phase2 with anchor_male). A non-anchor voice is zero-shot: its clip is the flow prompt and the LM prompt, and its
+speaker token is used when the LM was trained on that speaker.
+
+  python infer.py --text "<eng>this is dynamic accent,</eng> <pcm>e dey sound nice abi?</pcm>" --version phase3 --out out.wav
+  python infer.py --text "..." --compare --out cmp.wav       # -> cmp_phase2.wav and cmp_phase3.wav, same text, seed, voice
+  python infer.py --list_voices
+  python infer.py --auto_tag --tag_only --text "Omo, this traffic no be here o. Ẹ ṣé púpọ̀."   # inspect / edit the tags
+
+Text: wrap each language span in <eng>, <pcm>, <yor>, <ibo> or <hau> tags. Untagged text is treated as English, unless
+--auto_tag (automatic word/span language ID, see tagger.py), --force_lang or --lang_map is given.
+`--pause_mode clause` / `--speed` (prosody v2: fixed gaps + WSOLA) are kept for phase2 only as an experiment; a
+listening test preferred the model's own pauses.
 """
-import argparse, inspect, json, os, re, sys, unicodedata
+
+import argparse, copy, inspect, json, os, re, sys, unicodedata
 
 import numpy as np
 import soundfile as sf
 import librosa
 import torch
 import torch.nn as nn
+import contextlib
 
 HERE = os.path.dirname(os.path.abspath(__file__))
 LANGS = ["eng", "ibo", "hau", "yor", "pcm"]
 L2I = {l: i for i, l in enumerate(LANGS)}
 ANCHOR = "waxal_yor_4"
+DEV = torch.device("cuda" if torch.cuda.is_available() else "cpu")   # set_device() / --device to override
 DEFAULT_SPEED = 1.0    # user listening test 2026-10-21: original model pauses preferred; clause/1.05 kept as experimental
 
 
+def set_device(d):
+    global DEV
+    DEV = torch.device(d if d != "auto" else ("cuda" if torch.cuda.is_available() else "cpu"))
+    return DEV
+
+
+def autocast(dtype):
+    if DEV.type == "cuda" and dtype is not None:
+        return torch.autocast("cuda", dtype=dtype)
+    return contextlib.nullcontext()            # CPU: plain fp32
+
+
 # ------------------------------------------------------------------ text front-end (tags -> spans)
 _TAG = re.compile(r"<(eng|ibo|hau|yor|pcm)>(.*?)</\1>", re.S)
 _SENT = re.compile(r"(?<=[.?!])\s+")
@@ -214,6 +237,11 @@ def _sf_load_wav(wav, target_sr, min_sr=16000):
 
 def load_cosy(model_dir, cosyvoice_dir):
     sys.path[:0] = [cosyvoice_dir, os.path.join(cosyvoice_dir, "third_party", "Matcha-TTS")]
+    try:                                   # pyworld is only needed for CosyVoice training (F0 features); it often fails
+        import pyworld  # noqa: F401       # to build, so inference runs without it
+    except ImportError:
+        import types
+        sys.modules["pyworld"] = types.ModuleType("pyworld")
     import cosyvoice.utils.file_utils as fu
     fu.load_wav = _sf_load_wav
     import cosyvoice.cli.frontend as cf
@@ -231,16 +259,16 @@ def prompt_feats(cosy, wav_path):
     n = min(feat.shape[1] // 2, tokn.shape[1])
     feat, tokn = feat[:, :2 * n], tokn[:, :n]
     emb = fe._extract_spk_embedding(wav_path)
-    return tokn.cuda(), feat.cuda(), emb.cuda()
+    return tokn.to(DEV), feat.to(DEV), emb.to(DEV)
 
 
 @torch.no_grad()
 def token2wav(cosy, tokens, ptok, pfeat, emb):
     flow = cosy.model.flow
-    t = torch.tensor([tokens], dtype=torch.int32).cuda()
-    kw = dict(token=t, token_len=torch.tensor([t.shape[1]], dtype=torch.int32).cuda(),
-              prompt_token=ptok, prompt_token_len=torch.tensor([ptok.shape[1]], dtype=torch.int32).cuda(),
-              prompt_feat=pfeat, prompt_feat_len=torch.tensor([pfeat.shape[1]], dtype=torch.int32).cuda(),
+    t = torch.tensor([tokens], dtype=torch.int32).to(DEV)
+    kw = dict(token=t, token_len=torch.tensor([t.shape[1]], dtype=torch.int32).to(DEV),
+              prompt_token=ptok, prompt_token_len=torch.tensor([ptok.shape[1]], dtype=torch.int32).to(DEV),
+              prompt_feat=pfeat, prompt_feat_len=torch.tensor([pfeat.shape[1]], dtype=torch.int32).to(DEV),
               embedding=emb)
     sig = inspect.signature(flow.inference).parameters
     if "streaming" in sig:
@@ -248,43 +276,55 @@ def token2wav(cosy, tokens, ptok, pfeat, emb):
     if "finalize" in sig:
         kw["finalize"] = True
     if "flow_cache" in sig:
-        kw["flow_cache"] = torch.zeros(1, 80, 0, 2).cuda()
+        kw["flow_cache"] = torch.zeros(1, 80, 0, 2).to(DEV)
     mel = flow.inference(**kw)
     mel = mel[0] if isinstance(mel, tuple) else mel
     hs = inspect.signature(cosy.model.hift.inference).parameters
-    w = cosy.model.hift.inference(speech_feat=mel, **({"cache_source": torch.zeros(1, 1, 0).cuda()} if "cache_source" in hs else {}))
+    w = cosy.model.hift.inference(speech_feat=mel, **({"cache_source": torch.zeros(1, 1, 0).to(DEV)} if "cache_source" in hs else {}))
     w = w[0] if isinstance(w, tuple) else w
     return w.squeeze().float().cpu().numpy()
 
 
 # ------------------------------------------------------------------ Dynamic Accent LM (Qwen2LM + language + speaker embeddings)
+# S3 tokens that fill silent frames in the anchor's recordings (Phase 3 Stage 0); used by the silence-aware sampler
+SIL_TOKENS = {4137, 4218, 1950, 2031, 3894, 6324, 1707, 3677, 1701, 6405, 1947, 3645, 1944, 1704, 3891, 4134, 3680, 5838, 3651}
+RATES = [None, "very slow", "slow", "normal", "fast", "very fast"]
+
+
 class DALM2(nn.Module):
     def __init__(self, qlm, n_spk, d=896):
         super().__init__()
         self.q = qlm
         self.lang_emb = nn.Embedding(len(LANGS), d)
         self.spk_emb = nn.Embedding(n_spk, d)
+        self.rate_emb = nn.Embedding(len(RATES), d, padding_idx=0)   # phase3 speed instruct; zero = none (phase2)
+        nn.init.zeros_(self.rate_emb.weight)
         self.adv = nn.Sequential(nn.Linear(d, 256), nn.ReLU(), nn.Linear(256, n_spk))  # training-only head
         self.V = qlm.speech_token_size
 
     def embed_text(self, ids, langs):
         return self.q.llm.model.model.embed_tokens(ids) + self.lang_emb(langs)
 
-    def build2(self, ids, langs, spk, pids=None, plangs=None, psp=None):
+    def build2(self, ids, langs, spk, rate=0, pids=None, plangs=None, psp=None):
         sos = self.q.llm_embedding.weight[self.q.sos][None]
         task = self.q.llm_embedding.weight[self.q.task_id][None]
         if pids is not None:
             ids = torch.cat([pids, ids]); langs = torch.cat([plangs, langs])
-        parts = [sos, self.embed_text(ids, langs), self.spk_emb(spk)[None], task]
+        sv = self.spk_emb(spk)
+        if rate:
+            sv = sv + self.rate_emb(torch.tensor(rate, device=sv.device))
+        parts = [sos, self.embed_text(ids, langs), sv[None], task]
         if psp is not None:
             parts.append(self.q.speech_embedding(psp))
         return torch.cat(parts, 0)
 
     @torch.no_grad()
     def generate2(self, ids, langs, spk, prompt=None, top_p=0.8, top_k=25, win=10, tau_r=0.1, seed=0,
-                  min_ratio=2.0, max_ratio=20.0):
+                  min_ratio=2.0, max_ratio=20.0, sampler="ras", rate=0, sil_cap=25):
+        """sampler 'ras' = Phase 2 repetition-aware sampling; 'silexcap' = silence tokens exempt from the repetition
+        check while the current silence run is shorter than sil_cap frames (25 = 1.0 s), then the normal check."""
         g = torch.Generator(device=ids.device).manual_seed(seed)
-        x = self.build2(ids, langs, spk, *(prompt if prompt else (None, None, None)))[None]
+        x = self.build2(ids, langs, spk, rate, *(prompt if prompt else (None, None, None)))[None]
         n = len(ids)
         min_len, max_len = int(n * min_ratio), int(n * max_ratio)
         past, out, inp = None, [], x
@@ -301,7 +341,15 @@ class DALM2(nn.Module):
             keep[top_k:] = False
             cp, ci = sp[keep], si[keep]
             tokid = int(ci[torch.multinomial(cp / cp.sum(), 1, generator=g)])
-            if out and sum(1 for t in out[-win:] if t == tokid) >= win * tau_r and tokid != self.V:  # RAS
+            exempt = False
+            if sampler == "silexcap" and tokid in SIL_TOKENS:
+                run = 0
+                for t in reversed(out):
+                    if t not in SIL_TOKENS:
+                        break
+                    run += 1
+                exempt = run < sil_cap
+            if not exempt and out and sum(1 for t in out[-win:] if t == tokid) >= win * tau_r and tokid != self.V:  # RAS
                 tokid = int(torch.multinomial(p, 1, generator=g))
             if tokid >= self.V:
                 break
@@ -310,40 +358,128 @@ class DALM2(nn.Module):
         return out
 
 
+SETTINGS_PATH = os.path.join(HERE, "settings.json")
+SETTINGS = json.load(open(SETTINGS_PATH)) if os.path.exists(SETTINGS_PATH) else {}
+VERSIONS = {k: {kk: vv for kk, vv in v.items() if kk != "desc"} for k, v in SETTINGS.get("versions", {}).items()} or {
+    "phase2": dict(lm="checkpoints/phase2/p2_lm_B/best.pt", flow="checkpoints/phase2/p2_flow_sft/best.pt", sampler="ras",
+                   one_pass=False, postprocess=True, rate=None),
+    "phase3": dict(lm="checkpoints/phase3/p3_lm_full/best.pt", flow="checkpoints/phase2/p2_flow_sft/best.pt",
+                   sampler="silexcap", one_pass=True, postprocess=False, rate="normal"),
+}
+RECOMMENDED = SETTINGS.get("recommended", {})
+
+
+def recommendation(version, voice):
+    """one-line hint when a version/voice pairing is not the recommended one (None when it is)."""
+    rec = RECOMMENDED.get(version)
+    if rec and voice not in rec:
+        return f"note: {version} is recommended with {' or '.join(rec)} ({RECOMMENDED.get('note', '')})"
+    return None
+
+
+def version_cfg(name):
+    cfg = dict(VERSIONS[name])
+    cj = os.path.join(HERE, os.path.dirname(cfg["lm"]), "config.json")
+    if os.path.exists(cj):                      # a checkpoint may override flow / rate / sampler
+        cfg.update({k: v for k, v in json.load(open(cj)).items() if k in cfg})
+    cfg["lm"] = os.path.join(HERE, cfg["lm"]); cfg["flow"] = os.path.join(HERE, cfg["flow"])
+    return cfg
+
+
+def word_chunks(text, max_words=60):
+    """one pass per <= max_words, split only at sentence ends (a whole paragraph usually fits in one pass)."""
+    sents = split_sentences(text)
+    out, cur, n = [], [], 0
+    for s in sents:
+        w = len(re.sub(r"<[^>]+>", " ", s).split())
+        if cur and n + w > max_words:
+            out.append(" ".join(cur)); cur, n = [], 0
+        cur.append(s); n += w
+    if cur:
+        out.append(" ".join(cur))
+    return out or [text]
+
+
 class DynamicAccent:
-    def __init__(self, model_dir, cosyvoice_dir, lm_ckpt, flow_ckpt, prompts_json, dtype="auto", sent_pause=0.6, cap=0.6):
+    def __init__(self, model_dir, cosyvoice_dir, lm_ckpt, flow_ckpt, prompts_json, dtype="auto", sent_pause=0.6, cap=0.6,
+                 sampler="ras", one_pass=False, postprocess=True, rate=None, cosy=None, tok=None, copy_llm=False):
         from transformers import AutoTokenizer
-        self.tok = AutoTokenizer.from_pretrained(os.path.join(model_dir, "CosyVoice-BlankEN"))
-        self.cosy = load_cosy(model_dir, cosyvoice_dir)
+        self.tok = tok or AutoTokenizer.from_pretrained(os.path.join(model_dir, "CosyVoice-BlankEN"))
+        self.cosy = cosy or load_cosy(model_dir, cosyvoice_dir)
         self.sr = self.cosy.sample_rate
-        self.cosy.model.flow.load_state_dict(torch.load(flow_ckpt, map_location="cuda")["flow"])
-        ck = torch.load(lm_ckpt, map_location="cuda")
+        self.flow_sd = torch.load(flow_ckpt, map_location=DEV, weights_only=False)["flow"]
+        self.cosy.model.flow.load_state_dict(self.flow_sd)
+        ck = torch.load(lm_ckpt, map_location=DEV, weights_only=False)
         self.spk2id = json.load(open(os.path.join(os.path.dirname(lm_ckpt), "spk2id.json")))
-        self.m = DALM2(self.cosy.model.llm, ck["n_spk"]).cuda().eval()
-        self.m.load_state_dict(ck["model"])
-        if dtype == "auto":
+        # two versions resident at once (app.py) must not share LM weights: copy_llm=True gives each its own copy.
+        # The CLI runs versions one after the other and reloads the full LM state in place (saves ~2 GB RAM).
+        qlm = copy.deepcopy(self.cosy.model.llm) if copy_llm else self.cosy.model.llm
+        self.m = DALM2(qlm, ck["n_spk"]).to(DEV).eval()
+        miss = self.m.load_state_dict(ck["model"], strict=False)
+        assert not [k for k in miss.missing_keys if not k.startswith("rate_emb")], miss
+        del ck
+        if DEV.type != "cuda":
+            dtype = "fp32"
+        elif dtype == "auto":
             dtype = "bf16" if torch.cuda.is_bf16_supported() else "fp16"
-        self.dtype = torch.bfloat16 if dtype == "bf16" else torch.float16
+        self.dtype = {"bf16": torch.bfloat16, "fp16": torch.float16, "fp32": None}[dtype]
         self.sent_pause, self.cap = sent_pause, cap
+        self.sampler, self.one_pass, self.postprocess = sampler, one_pass, postprocess
+        self.rate = RATES.index(rate) if rate in RATES else 0
         meta = json.load(open(prompts_json, encoding="utf-8"))
         root = os.path.dirname(os.path.dirname(os.path.abspath(prompts_json)))
         self.prompts = []
         for p in meta["prompts"]:
             ids, langs = text_tokens(self.tok, p["tagged"])
-            lmp = (torch.tensor(ids).cuda(), torch.tensor(langs).cuda(), torch.tensor(p["s3_tokens"], dtype=torch.int64).cuda())
-            self.prompts.append(dict(id=p["id"], lm=lmp, flow=prompt_feats(self.cosy, os.path.join(root, p["wav"]))))
-
-    def one(self, tagged_text, seed, prompt_k, use_lm_prompt=True):
+            lmp = (torch.tensor(ids).to(DEV), torch.tensor(langs).to(DEV), torch.tensor(p["s3_tokens"], dtype=torch.int64).to(DEV))
+            self.prompts.append(dict(id=p["id"], lm=lmp, flow=prompt_feats(self.cosy, os.path.join(root, p["wav"])), spk=ANCHOR))
+        self.voice = None
+
+    def use_flow(self):
+        self.cosy.model.flow.load_state_dict(self.flow_sd)       # two versions may share one CosyVoice2 instance
+
+    def set_voice(self, voice, voices_json=None):
+        """'anchor' or an id from voices.json: its clip is the flow + LM prompt (zero-shot)."""
+        meta = json.load(open(voices_json, encoding="utf-8")) if voices_json else {"voices": {}}
+        voice = meta.get("aliases", {}).get(voice, voice)
+        if voice in (None, "anchor", "anchor_male"):
+            self.voice = None; return
+        if voice not in meta["voices"]:
+            raise SystemExit(f"unknown voice '{voice}'; available: {', '.join(meta['voices'])}")
+        V = meta["voices"][voice]
+        root = os.path.dirname(os.path.dirname(os.path.abspath(voices_json)))
+        ptok, pfeat, pemb = prompt_feats(self.cosy, os.path.join(root, V["wav"]))
+        ids, langs = text_tokens(self.tok, V["tagged"])
+        lmp = (torch.tensor(ids).to(DEV), torch.tensor(langs).to(DEV), ptok[0].long())
+        self.voice = dict(id=voice, lm=lmp, flow=(ptok, pfeat, pemb), spk=V["spk"] if V["spk"] in self.spk2id else ANCHOR)
+
+    def _prompt(self, prompt_k):
+        return self.voice or self.prompts[prompt_k % len(self.prompts)]
+
+    def one(self, tagged_text, seed, prompt_k, use_lm_prompt=True, lm_prompt=None):
         ids, langs = text_tokens(self.tok, tagged_text)
-        P = self.prompts[prompt_k % len(self.prompts)]
-        with torch.autocast("cuda", dtype=self.dtype):
-            sp = self.m.generate2(torch.tensor(ids).cuda(), torch.tensor(langs).cuda(),
-                                  torch.tensor(self.spk2id[ANCHOR]).cuda(), prompt=P["lm"] if use_lm_prompt else None, seed=seed)
+        P = self._prompt(prompt_k)
+        with autocast(self.dtype):
+            sp = self.m.generate2(torch.tensor(ids).to(DEV), torch.tensor(langs).to(DEV), torch.tensor(self.spk2id[P["spk"]]).to(DEV),
+                                  prompt=(lm_prompt or P["lm"]) if use_lm_prompt else None, seed=seed, sampler=self.sampler,
+                                  rate=self.rate, max_ratio=30.0 if self.one_pass else 20.0)
         if len(sp) < 3:
             sp = sp + [0] * 3
+        self.last_tokens = sp
         return token2wav(self.cosy, sp, *P["flow"])
 
     def say(self, tagged_text, seed=0, prompt_k=0, stitch=True, pause_mode="model", speed=1.0, return_info=False):
+        self.use_flow()
+        if self.one_pass:                      # phase3: the model makes every pause; nothing is trimmed or inserted
+            ws, prev = [], None
+            for i, ch in enumerate(word_chunks(tagged_text)):
+                w = self.one(ch, seed + 1000 * i, prompt_k, lm_prompt=prev)
+                ws.append(w)
+                ids, langs = text_tokens(self.tok, ch)       # next chunk continues from this one (model sets the gap)
+                if len(self.last_tokens) <= 400:
+                    prev = (torch.tensor(ids).to(DEV), torch.tensor(langs).to(DEV), torch.tensor(self.last_tokens).to(DEV))
+            w = np.concatenate(ws)
+            return (w, None) if return_info else w
         if pause_mode == "clause":
             return self.say_clause(tagged_text, seed, prompt_k, speed, return_info)
         w = self.say_model(tagged_text, seed, prompt_k, stitch)
@@ -359,11 +495,11 @@ class DynamicAccent:
             w = cap_internal(trim_tight(w, self.sr), self.sr, self.cap)
             w = fade(time_stretch(w, self.sr, speed), self.sr)
             if i == 0:
-                out.append(np.zeros(int(0.12 * self.sr), np.float32))       # same 0.12 s lead as model mode
+                out.append(np.zeros(int(0.12 * self.sr), np.float32))
             out.append(w)
             if b is not None:
                 g = GAPS[b] + rng.uniform(-JITTER[b], JITTER[b])
-                out.append(np.zeros(max(0, int((g - 2 * lead) * self.sr)), np.float32))  # 30 ms kept each side
+                out.append(np.zeros(max(0, int((g - 2 * lead) * self.sr)), np.float32))
             else:
                 out.append(np.zeros(int(0.15 * self.sr), np.float32))
             info.append(dict(lang=l, text=t, boundary=b))
@@ -384,34 +520,111 @@ class DynamicAccent:
         return np.concatenate(out)
 
 
+def build(version, a, cosy=None, tok=None, copy_llm=False):
+    cfg = version_cfg(version)
+    lm = a.lm or cfg["lm"]; flow = a.flow or cfg["flow"]
+    S = DynamicAccent(a.model_dir, a.cosyvoice_dir, lm, flow, a.prompts, a.dtype, sampler=cfg["sampler"],
+                      one_pass=cfg["one_pass"] and not a.stitch, postprocess=cfg["postprocess"],
+                      rate=a.rate if a.rate != "default" else cfg["rate"], cosy=cosy, tok=tok, copy_llm=copy_llm)
+    S.set_voice(a.voice, a.voices)
+    return S
+
+
+def list_voices(voices_json):
+    V = json.load(open(voices_json, encoding="utf-8"))["voices"] if os.path.exists(voices_json) else {}
+    print("voices (--voice):")
+    for k, v in V.items():
+        print(f"  {k:20s} {v.get('desc', '')}")
+    print("\nversions (--version):")
+    for k, v in SETTINGS.get("versions", {}).items():
+        print(f"  {k:20s} {v.get('desc', '')}")
+    if RECOMMENDED:
+        print("\nrecommended pairing: phase3 -> " + ", ".join(RECOMMENDED.get("phase3", [])) +
+              "; phase2 -> " + ", ".join(RECOMMENDED.get("phase2", [])))
+        print("  " + RECOMMENDED.get("note", ""))
+
+
+def prepare_text(a):
+    """apply --auto_tag / --force_lang / --lang_map; returns the tagged text that will be synthesised."""
+    from tagger import Tagger, parse_map, TAGGED
+    text = a.text
+    if a.force_lang:
+        from tagger import strip_tags
+        return f"<{a.force_lang}>{strip_tags(text)}</{a.force_lang}>"
+    if a.auto_tag or a.lang_map:
+        if TAGGED.search(text) and not a.auto_tag:
+            return text
+        from tagger import strip_tags
+        T = Tagger(None if a.lid == "none" else a.lid, parse_map(a.lang_map))
+        return T.tag(strip_tags(text))
+    return text
+
+
 def main():
     ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
-    ap.add_argument("--text", required=True, help="tagged text, e.g. '<eng>Good morning,</eng> <pcm>how body?</pcm>'")
+    ap.add_argument("--text", help="text; tagged ('<eng>Good morning,</eng> <pcm>how body?</pcm>') or plain with --auto_tag")
+    ap.add_argument("--version", default=SETTINGS.get("default_version", "phase3"), choices=sorted(VERSIONS),
+                    help="model version (default phase3)")
+    ap.add_argument("--compare", action="store_true",
+                    help="render the same text, seed and voice with phase2 AND phase3 -> <out>_phase2.wav, <out>_phase3.wav")
+    ap.add_argument("--voice", default=SETTINGS.get("default_voice", "anchor_male"),
+                    help="anchor_male | waxal_ibo_5_female | waxal_yor_5_female (see --list_voices)")
+    ap.add_argument("--list_voices", action="store_true", help="list voices, versions and the recommended pairing")
+    ap.add_argument("--rate", default="default", choices=["default", "none", "very slow", "slow", "normal", "fast", "very fast"],
+                    help="phase3 speed instruction to the LM (default: the version's setting, 'normal'; 'none' = no instruction)")
+    # code-switching control
+    ap.add_argument("--auto_tag", action="store_true",
+                    help="tag untagged mixed text automatically (word/span language ID), print it, then synthesise")
+    ap.add_argument("--tag_only", action="store_true", help="with --auto_tag: print the tagged text and exit (edit it, then pass it back as --text)")
+    ap.add_argument("--force_lang", choices=LANGS, help="read the whole text in one language (tags are ignored)")
+    ap.add_argument("--lang_map", default="", help='per-word language overrides for the tagger, e.g. "omo=pcm,sha=pcm"')
+    ap.add_argument("--lid", default="none", choices=["none", "glotlid"],
+                    help="optional LID model for --auto_tag (glotlid: cis-lmu/glotlid, Apache-2.0, 1.7 GB, pip install fasttext)")
     ap.add_argument("--seed", type=int, default=0)
     ap.add_argument("--out", default="out.wav")
-    ap.add_argument("--prompt_clip", type=int, default=0, help="anchor prompt clip 0-3 (final system: 0)")
+    ap.add_argument("--prompt_clip", type=int, default=0, help="anchor prompt clip 0-3 (final systems: 0)")
     ap.add_argument("--model_dir", default=os.environ.get("COSYVOICE2_DIR", os.path.join(HERE, "pretrained", "CosyVoice2-0.5B")),
-                    help="stock FunAudioLLM/CosyVoice2-0.5B snapshot")
+                    help="stock FunAudioLLM/CosyVoice2-0.5B snapshot (downloaded on first run)")
     ap.add_argument("--cosyvoice_dir", default=os.environ.get("COSYVOICE_REPO", os.path.join(HERE, "CosyVoice")),
                     help="clone of github.com/FunAudioLLM/CosyVoice (with submodules)")
-    ap.add_argument("--lm", default=os.path.join(HERE, "checkpoints", "p2_lm_B", "best.pt"))
-    ap.add_argument("--flow", default=os.path.join(HERE, "checkpoints", "p2_flow_sft", "best.pt"))
+    ap.add_argument("--lm", default="", help="override the version's LM checkpoint")
+    ap.add_argument("--flow", default="", help="override the version's flow checkpoint")
     ap.add_argument("--prompts", default=os.path.join(HERE, "prompts", "prompts.json"))
-    ap.add_argument("--dtype", default="auto", choices=["auto", "bf16", "fp16"],
-                    help="LM autocast dtype. Samples were made with bf16; T4 has no bf16, so auto picks fp16 there")
-    ap.add_argument("--no_stitch", action="store_true", help="generate multi-sentence text in one pass")
-    ap.add_argument("--pause_mode", default="model", choices=["clause", "model"],
-                    help="clause: clause-level chunks joined with anchor-matched gaps (default); model: Phase 2 final behaviour")
-    ap.add_argument("--speed", type=float, default=DEFAULT_SPEED,
-                    help="pitch-preserving tempo factor applied to speech (gaps are kept as set). 1.0 = Phase 2 final pace")
+    ap.add_argument("--voices", default=os.path.join(HERE, "prompts", "voices.json"))
+    ap.add_argument("--dtype", default="auto", choices=["auto", "bf16", "fp16", "fp32"],
+                    help="LM autocast dtype on GPU (samples: bf16; auto picks fp16 on GPUs without bf16). CPU always runs fp32")
+    ap.add_argument("--stitch", action="store_true", help="phase3: generate sentence by sentence instead of one pass")
+    ap.add_argument("--no_stitch", action="store_true", help="phase2: generate multi-sentence text in one pass")
+    ap.add_argument("--pause_mode", default="model", choices=["clause", "model"], help="phase2 only; clause = experimental prosody v2 (not recommended)")
+    ap.add_argument("--speed", type=float, default=DEFAULT_SPEED, help="phase2 only; pitch-preserving tempo factor (1.0 = as trained; not recommended)")
     a = ap.parse_args()
+    if a.list_voices:
+        list_voices(a.voices); return
+    if not a.text:
+        ap.error("--text is required")
+    text = prepare_text(a)
+    if text != a.text or a.auto_tag:
+        print(f"[tagged] {text}")
+    if a.tag_only:
+        return
     if not os.path.isdir(a.model_dir):
         from huggingface_hub import snapshot_download
         snapshot_download("FunAudioLLM/CosyVoice2-0.5B", local_dir=a.model_dir)
-    S = DynamicAccent(a.model_dir, a.cosyvoice_dir, a.lm, a.flow, a.prompts, a.dtype)
-    w = S.say(a.text, seed=a.seed, prompt_k=a.prompt_clip, stitch=not a.no_stitch, pause_mode=a.pause_mode, speed=a.speed)
-    sf.write(a.out, w, S.sr)
-    print(f"wrote {a.out} ({len(w) / S.sr:.2f} s, {S.sr} Hz)")
+    versions = ["phase2", "phase3"] if a.compare else [a.version]
+    cosy = tok = None
+    base, ext = os.path.splitext(a.out)
+    for v in versions:
+        hint = recommendation(v, json.load(open(a.voices)).get("aliases", {}).get(a.voice, a.voice) if os.path.exists(a.voices) else a.voice)
+        if hint:
+            print(hint)
+        S = build(v, a, cosy, tok)
+        cosy, tok = S.cosy, S.tok
+        w = S.say(text, seed=a.seed, prompt_k=a.prompt_clip, stitch=not a.no_stitch, pause_mode=a.pause_mode,
+                  speed=a.speed if v == "phase2" else 1.0)
+        out = f"{base}_{v}{ext or '.wav'}" if a.compare else a.out
+        sf.write(out, w, S.sr)
+        print(f"[{v}] wrote {out} ({len(w) / S.sr:.2f} s, {S.sr} Hz, voice={a.voice}, device={DEV.type})")
+        del S.m; torch.cuda.empty_cache()
 
 
 if __name__ == "__main__":

prompts/voice_waxal_ibo_5.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d537c83dbe167ccf0b43f03304e456fb05e59dfda6b4e359d941900ef629924f
+size 318044

prompts/voice_waxal_yor_5.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ce75b93cf5166d0dc2f53cc943330b4d304f2e060f831d38deb2e664fff49ca5
+size 182924

prompts/voices.json

@@ -0,0 +1,41 @@
+{
+ "note": "zero-shot voices: the clip is the flow + LM prompt; the speaker token is used when the LM saw that speaker",
+ "voices": {
+  "anchor_male": {
+   "wav": null,
+   "lang": "yor",
+   "spk": "waxal_yor_4",
+   "gender": "male",
+   "source": "WAXAL Media Trust TTS, Yoruba speaker 4 (prompts/clip0..3)",
+   "licence": "CC-BY-SA-4.0 (Media Trust TTS subset)",
+   "desc": "male, WAXAL Yoruba speaker 4: the anchor voice both versions were trained around (uses prompts/prompts.json clip 0)"
+  },
+  "waxal_ibo_5_female": {
+   "wav": "prompts/voice_waxal_ibo_5.wav",
+   "tagged": "<ibo>I nwere ike itinye mmemme ọbụla n'izu a ma gị atụla ụjọ maka oké mmiri ozuzo.</ibo>",
+   "text": "I nwere ike itinye mmemme ọbụla n'izu a ma gị atụla ụjọ maka oké mmiri ozuzo.",
+   "lang": "ibo",
+   "spk": "waxal_ibo_5",
+   "gender": "female",
+   "source_clip": "waxal_ibo_test_ibo_502_c003909",
+   "licence": "CC-BY-SA-4.0 (Media Trust TTS subset)",
+   "desc": "female native Igbo speaker (WAXAL ibo speaker 5); prompt clip waxal_ibo_test_ibo_502_c003909 (6.625 s)"
+  },
+  "waxal_yor_5_female": {
+   "wav": "prompts/voice_waxal_yor_5.wav",
+   "tagged": "<yor>Nínú oṣù Ọpẹ́ ni ó ra mọ́tò tí wọ́n jí gbé.</yor>",
+   "text": "Nínú oṣù Ọpẹ́ ni ó ra mọ́tò tí wọ́n jí gbé.",
+   "lang": "yor",
+   "spk": "waxal_yor_5",
+   "gender": "female",
+   "source_clip": "waxal_yor_test_yor_686",
+   "licence": "CC-BY-SA-4.0 (Media Trust TTS subset)",
+   "desc": "female native Yoruba speaker (WAXAL yor speaker 5); prompt clip waxal_yor_test_yor_686 (3.81 s)"
+  }
+ },
+ "aliases": {
+  "anchor": "anchor_male",
+  "waxal_ibo_5": "waxal_ibo_5_female",
+  "waxal_yor_5": "waxal_yor_5_female"
+ }
+}

requirements.txt

@@ -21,7 +21,7 @@ inflect>=7.3
 x-transformers>=2.11
 openai-whisper>=20231117
 einops>=0.8
-pyworld>=0.3.4
+# pyworld>=0.3.4   # CosyVoice training only; infer.py runs without it
 pyarrow>=18
 wget==3.2
 gdown>=5.1
@@ -31,3 +31,5 @@ protobuf>=4.25
 pydantic>=2.7
 wetext==0.0.4
 audiotsm==0.1.2   # WSOLA time-stretch for --speed (falls back to librosa if missing)
+gradio>=4.44,<6   # app.py
+# optional, for --lid glotlid / the GlotLID checkbox:  fasttext (or fasttext-wheel)

samples/phase3/phase2/para_hau_eng_pcm.mp3

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cd687cc31eab7f893afd9e8ab1e64a3ce2093c7331ef6a70cd0a5f9a3f33bd72
+size 143876

samples/phase3/phase2/para_ibo_pcm_eng.mp3

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:afae973f5a0fcbd50931b8dfc8f42d48c91b17f18a584a6558c8609af0e37f41
+size 122636

samples/phase3/phase2/para_yor_eng.mp3

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:03d0f2bce66da138cbe6e0fe2d38392e638122b82be82dfcb1ba83b3f140d811
+size 123380

samples/phase3/phase3/para_hau_eng_pcm.mp3

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ee8e7b3b1011ef0e6fd6900acae4d7ddc62cbb830b0401492f6a5e40a84b19fa
+size 132620

samples/phase3/phase3/para_ibo_pcm_eng.mp3

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be778e8be5554b933e8fcaa5d9deac5150b2defd5fca9351bc3ae483697ed5b4
+size 128036

samples/phase3/phase3/para_yor_eng.mp3

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2546d9ddde91b773447148cf767f19252bad7efc0914b494e04ee1cf8a23cd99
+size 123884

samples/phase3/phase3_female/para_ibo_pcm_eng__waxal_ibo_5.mp3

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eacca63c857ff58afcce75297eeb4c7e7f9deeb90ebc2dd212e57a026a53121a
+size 119228

samples/phase3/phase3_female/para_yor_eng__waxal_yor_5.mp3

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0f6b90c343f255c8ffaab0ea92811928dee7b8b985fcd14973c2cbec40f80429
+size 119060

settings.json

@@ -0,0 +1,47 @@
+{
+ "default_version": "phase3",
+ "default_voice": "anchor_male",
+ "versions": {
+  "phase2": {
+   "lm": "checkpoints/phase2/p2_lm_B/best.pt",
+   "flow": "checkpoints/phase2/p2_flow_sft/best.pt",
+   "sampler": "ras",
+   "one_pass": false,
+   "postprocess": true,
+   "rate": null,
+   "desc": "Phase 2 final system: LM p2_lm_B (~120 h, pause profile B) + anchor-SFT flow; sentence by sentence, internal silences capped at 0.6 s"
+  },
+  "phase3": {
+   "lm": "checkpoints/phase3/p3_lm_full/best.pt",
+   "flow": "checkpoints/phase2/p2_flow_sft/best.pt",
+   "sampler": "silexcap",
+   "one_pass": true,
+   "postprocess": false,
+   "rate": "normal",
+   "desc": "Phase 3: LM p3_lm_full (178 h re-cut, ~102 h Igbo, [pause]/[breath] + speed instruction) + the shared Phase 2 flow; one pass per paragraph, model-made pauses"
+  }
+ },
+ "voices": [
+  "anchor_male",
+  "waxal_ibo_5_female",
+  "waxal_yor_5_female"
+ ],
+ "recommended": {
+  "phase3": [
+   "waxal_ibo_5_female",
+   "waxal_yor_5_female"
+  ],
+  "phase2": [
+   "anchor_male"
+  ],
+  "note": "Listening verdict (model author, 2026-10-21): Phase 3 sounds best with the female voices; Phase 2 is fine for the male anchor. Use --compare to A/B before use."
+ },
+ "rates": [
+  "none",
+  "very slow",
+  "slow",
+  "normal",
+  "fast",
+  "very fast"
+ ]
+}

tagger.py

@@ -0,0 +1,348 @@
+#!/usr/bin/env python
+"""Dynamic Accent auto-tagger: word/span-level language ID for untagged mixed Nigerian text.
+
+Labels: eng (English), pcm (Nigerian Pidgin), ibo (Igbo), yor (Yoruba), hau (Hausa).
+Output is the tagged text infer.py expects, e.g.
+    "omo, I'm very excited"  ->  "<pcm>omo,</pcm> <eng>I'm very excited</eng>"
+
+How it decides (rule-based, no model needed):
+  1. letters: Yoruba-only (ẹ ṣ, tone-marked n/m), Igbo-only (ị ụ ṅ), Hausa-only (ɓ ɗ ƙ ƴ), shared ọ;
+  2. lexicons of frequent function words / greetings per language (diacritics folded for lookup);
+  3. Pidgin markers ("dey", "abi", "omo", "wetin", "abeg", "una", "no go", "e don", "know say", pronoun + "go" + verb ...):
+     a clause of plain-ASCII words that contains a Pidgin marker becomes Pidgin as a whole;
+  4. clause context: ambiguous words ("na", "o", "ka", "ya", "ma", "ni" ...) and unknown words take the language of the
+     clause they sit in, then of the neighbouring clause; plain ASCII with no evidence falls back to English.
+  5. optional: --lid glotlid adds clause-level scores from GlotLID v3 (cis-lmu/glotlid, fastText, Apache-2.0 + notices,
+     1.7 GB download, needs `pip install fasttext`), mainly to separate untagged Igbo/Yoruba/Hausa written without
+     diacritics from English.
+
+Overrides: lang_map {"omo": "pcm"} forces a word; force_lang renders everything in one language.
+Always print the tagged text and let the user fix it: the tagger is a convenience, not ground truth.
+
+    python tagger.py "Omo, this traffic no be here o. I go call you later."
+    python tagger.py --lang_map "omo=pcm,sha=pcm" "..."
+"""
+import argparse, json, re, sys, unicodedata
+
+LANGS = ["eng", "pcm", "ibo", "yor", "hau"]
+NG = ("ibo", "yor", "hau")
+TAGGED = re.compile(r"<(eng|ibo|hau|yor|pcm)>")
+
+
+def fold(w):
+    """lowercase, NFD, drop every combining mark (tones, dot below, dot above), map hooked letters."""
+    w = unicodedata.normalize("NFD", w.lower())
+    w = "".join(c for c in w if not unicodedata.combining(c))
+    return w.translate(str.maketrans({"ɓ": "b", "ɗ": "d", "ƙ": "k", "ƴ": "y", "’": "'", "ʼ": "'", "‘": "'"}))
+
+
+def _set(s):
+    return {fold(x) for x in s.split()}
+
+
+def _exact(s):
+    return {unicodedata.normalize("NFC", x.lower()) for x in s.split()}
+
+
+# --- lexicons (frequent, high-precision words; folded). Words shared across languages are listed in each.
+_RAW = {
+    "ibo": """
+        bụ nke ndị ihe ụlọ mmiri kedu ndewo nụ ụmụ ụmụnna nna nne anyị unu gị ha ọ ya ka ma
+        nwere ga ji dị ọma biko daalụ dalu ụtọ obi ike nri ego akwụkwọ ọrụ maka ebe otu ọzọ ugbu
+        a m n' na mere meela kwa kwesịrị ịga ịbịa bịa gaa nyere aka mmadụ obodo nwanne nwa nwoke nwanyị
+        ụbọchị echi taa ụnyaahụ abalị ututu ehihie mgbede oge afọ ọnwa izu ukwuu nnukwu ntakịrị
+        mana ọbụna ọbụla ruo site gbasara ebee onye gịnị olee ole kpọmkwem eziokwu nso anọ abụọ atọ
+        ise isii asaa asatọ itoolu iri mbụ zuru ezu ahụ ahụla ịhụ hụ chere chọrọ chọ mara amaghị
+        nọ nọrọ bịara gara kwuru kwuo sị siri esi na-esi na-eme eme mee eji ogologo ọkụ ụzọ ahịa ụka
+        ọnụ ụtụtụ ka ọ dị mma mma nnọọ kpọọ zụta re ree rie riri ụmụaka ezinụlọ chi chukwu ndụ
+        """,
+    "yor": """
+        ni ti si sí àti àwọn mo o wa wọn kò ṣé jẹ́ ọjọ́ ẹ káàbọ̀ ilé ìlú ọmọ púpọ̀ dáadáa bẹ́ẹ̀ ni
+        náà yìí kan mi wa yín rẹ̀ rẹ tirẹ̀ fún láti pé ń máa yóò ó a jẹ ṣe lọ wá rí fẹ́ mọ̀ sọ gbọ́
+        ẹ̀kọ́ iṣẹ́ owó ọjà òní ọ̀la àná ìpàdé padà ilé-ìwé ọjọ́ ajé ìsinmi alẹ́ àárọ̀ káàárọ̀ káàsán
+        káalẹ́ ẹ kú ṣé gan an pàdé ìyá bàbá ẹ̀gbọ́n àbúrò ọ̀rẹ́ ènìyàn èèyàn àgbà ọba ìjọba
+        nítorí ṣùgbọ́n tàbí bí kí kì ìgbà ibi ibo kílódé báwo kíni ta tani mélòó ọ̀kan méjì mẹ́ta
+        mẹ́rin márùn-ún ẹ̀ẹ̀kan dúpẹ́ ẹ ṣeun ẹ̀ jọ̀ọ́ jọ̀ọ́ oúnjẹ omi ilẹ̀ ojú ọwọ́ ẹsẹ̀
+        """,
+    "hau": """
+        da ba ya ta ka ki su mu na ne ce ina yaya barka kuke don haka za idan mai wannan wancan yana
+        tana suna muna kuna sannu lafiya kwana aiki gobe yau jiya ruwa sai amma kuma shi ita nan
+        wani wata abin zan zai zata zamu mutane gida kudi gode sosai tafi fita zuwa laima yamma safe
+        rana dare mako shekara wata yanzu kai ke ku mun sun ban bai bata sani so iya yi yin
+        farashin tumatur tashi kasuwa abinci ruwan sha babu akwai kamar saboda domin tare har tun
+        daga cikin waje gaba baya sama kasa me wane wace ina yaushe nawa daya biyu uku hudu biyar
+        """,
+    "pcm": """
+        dey abi omo wetin abeg una wahala sabi oga sef sha oya jare ehn ehen shey kpele pikin
+        chop comot waka belle palava wan wey dem don am siddon gist gbam yawa kampe sharp sharp
+        nawa wahala kolo japa ginger vex dash ogbonge padi paddy shine eye
+        """,
+    "eng": """
+        the a an and or but if then so because of to in on at for with from by about as into over
+        after before between through during without under again further here there when where why how
+        all any both each few more most other some such no nor not only own same than too very can
+        will just should now i me my we our you your he him his she her it its they them their what
+        which who whom this that these those am is are was were be been being have has had having do
+        does did doing would could might must shall may please thank thanks yes hello hi good morning
+        evening afternoon night today tomorrow yesterday week year time day people man woman child
+        children team meeting report school market road weather rain safe everyone everybody
+        new old big small great right left back home work job money food water house car family
+        friend friends know think say said tell told make made take took get got give gave go went
+        come came see saw look want need like love help call send sent finish start stop keep let put
+        read write talk speak discuss three two one four five six seven eight nine ten first second
+        last next soon late early ready sure also still really very much many lot thing things way
+        out up down off well done going coming doing nothing something everything anything model
+        """,
+}
+LEX = {l: _set(v) for l, v in _RAW.items()}          # folded (diacritics dropped) for lookup
+LEX_EXACT = {l: _exact(v) for l, v in _RAW.items()}  # as written
+_N_LEX = {}
+for _l, _v in LEX.items():
+    for _w in _v:
+        _N_LEX[_w] = _N_LEX.get(_w, 0) + 1          # >1 = word shared by several languages (na, o, ka, ya ...)
+# words that are Pidgin markers on their own (strong evidence)
+PCM_STRONG = _set("dey abi omo wetin abeg una wahala sabi sef oya jare shey pikin chop comot waka palava wan wey dem siddon vex japa ehn ehen")
+PCM_BIGRAMS = {("no", "go"), ("no", "be"), ("no", "fit"), ("no", "dey"), ("e", "don"), ("e", "dey"), ("e", "no"),
+               ("know", "say"), ("tell", "am"), ("well", "well"), ("how", "far"), ("how", "body"), ("na", "so"),
+               ("make", "we"), ("make", "i"), ("make", "una"), ("make", "dem"), ("i", "no"), ("we", "no"),
+               ("you", "no"), ("dem", "no"), ("no", "easy"), ("no", "forget"), ("no", "vex"), ("fit", "heavy"),
+               ("don", "come"), ("don", "go"), ("for", "inside"), ("una", "well"), ("e", "go"), ("na", "you"),
+               ("na", "me"), ("na", "im"), ("na", "dem"), ("na", "una"), ("give", "you"), ("lie", "give")}
+PRON = _set("i we you dem una e im him she he they")
+NOT_VERB_AFTER_GO = _set("to home there back out away on in up down off ahead")
+PCM_AMBIG_END = _set("o")          # clause-final particle "o" in an ASCII clause -> Pidgin evidence
+
+YOR_CH = set("ẹṣ")                   # after NFC lowercase; also e/s + U+0323
+IBO_CH = set("ịụṅ")
+HAU_CH = set("ɓɗƙƴ")
+WORD = re.compile(r"[^\W\d_](?:[^\W\d_]|['’ʼ\-\u0300-\u036f])*", re.U)
+CLAUSE_END = re.compile(r"[,;:.?!]")
+
+
+def char_scores(w):
+    s = dict.fromkeys(LANGS, 0.0)
+    nfd = unicodedata.normalize("NFD", w.lower())
+    base = None
+    for c in nfd:
+        if unicodedata.combining(c):
+            if c == "\u0323":                    # dot below
+                if base in "es":
+                    s["yor"] += 3
+                elif base in "iu":
+                    s["ibo"] += 3
+                elif base == "o":
+                    s["yor"] += 1; s["ibo"] += 1
+            elif c == "\u0307" and base == "n":
+                s["ibo"] += 3
+            elif c in "\u0300\u0301\u0304":       # tones
+                if base in "nm":
+                    s["yor"] += 2
+                else:
+                    s["yor"] += 0.8; s["ibo"] += 0.3
+            continue
+        base = c
+        if c in HAU_CH:
+            s["hau"] += 4
+    if not any(s.values()) and re.fullmatch(r"[a-z'\-]+", fold(w)):
+        s["ascii"] = 1
+    return s
+
+
+def word_scores(w):
+    s = char_scores(w)
+    ascii_ = s.pop("ascii", 0)
+    f = fold(w)
+    parts = [f] + (f.split("-") if "-" in f else [])
+    if f.startswith("n'") or re.match(r"^na-[a-z]", f):
+        s["ibo"] += 3
+    ex = unicodedata.normalize("NFC", w.lower())
+    for p in parts:
+        amb = 0.5 if _N_LEX.get(p, 0) > 1 else 1.0          # shared words: let the clause decide
+        for l in LANGS:
+            if p in LEX[l]:
+                w_ = 1.5 if (p == f and ex in LEX_EXACT[l]) or ascii_ and p in LEX_EXACT[l] else 0.7
+                s[l] += amb * (w_ if l != "pcm" else 1.0)
+    if f in PCM_STRONG and ascii_:                           # Pidgin markers are plain ASCII ("ọmọ" is Yoruba)
+        s["pcm"] += 4
+    # orthography hints (weak)
+    if ascii_:
+        if re.search(r"(th|wh|ck|tion|ing$|ould$|ght)", f):
+            s["eng"] += 1
+        if re.match(r"^(nw|kw|gw|mm|nn|nk|ng[a-z])", f) and f not in LEX["eng"]:
+            s["ibo"] += 0.7
+        if re.search(r"(ts|^ky|^'y|aa|ii|uu)", f) and f not in LEX["eng"]:
+            s["hau"] += 0.5
+    return s, bool(ascii_)
+
+
+def tokenize(text):
+    """-> list of (start, end, word) for words; everything else is glue."""
+    return [(m.start(), m.end(), m.group(0)) for m in WORD.finditer(text)]
+
+
+def clauses(text, toks):
+    """clause index for each word: a new clause starts after , ; : . ? !"""
+    ends = [m.start() for m in CLAUSE_END.finditer(text)]
+    out, k = [], 0
+    for a, b, w in toks:
+        while k < len(ends) and ends[k] < a:
+            k += 1
+        out.append(k)
+    return out
+
+
+def _norm(d):
+    t = sum(max(0.0, v) for v in d.values())
+    return {k: (max(0.0, v) / t if t else 0.0) for k, v in d.items()}
+
+
+class Tagger:
+    def __init__(self, lid=None, lang_map=None):
+        self.lang_map = {fold(k): v for k, v in (lang_map or {}).items()}
+        self.lid = None
+        if lid == "glotlid":
+            import fasttext
+            from huggingface_hub import hf_hub_download
+            self.lid = fasttext.load_model(hf_hub_download("cis-lmu/glotlid", "model.bin"))
+
+    def _lid_scores(self, text):
+        labs, ps = self.lid.predict(text.replace("\n", " "), k=20)
+        m = {"__label__eng_Latn": "eng", "__label__pcm_Latn": "pcm", "__label__ibo_Latn": "ibo",
+             "__label__yor_Latn": "yor", "__label__hau_Latn": "hau"}
+        s = dict.fromkeys(LANGS, 0.0)
+        for l, p in zip(labs, ps):
+            if l in m:
+                s[m[l]] += float(p)
+        return s
+
+    def word_langs(self, text):
+        toks = tokenize(text)
+        if not toks:
+            return toks, []
+        cl = clauses(text, toks)
+        n = len(toks)
+        sc, asc = zip(*[word_scores(w) for _, _, w in toks])
+        sc = [dict(s) for s in sc]
+        f = [fold(w) for _, _, w in toks]
+        # Pidgin bigrams and "pronoun go verb"
+        for i in range(n - 1):
+            if cl[i] != cl[i + 1]:
+                continue
+            if (f[i], f[i + 1]) in PCM_BIGRAMS:
+                sc[i]["pcm"] += 2.5; sc[i + 1]["pcm"] += 2.5
+            if i + 2 < n and f[i] in PRON and f[i + 1] == "go" and cl[i + 2] == cl[i] and f[i + 2] not in NOT_VERB_AFTER_GO \
+                    and f[i + 2] in LEX["eng"] | LEX["pcm"]:
+                for j in (i, i + 1, i + 2):
+                    sc[j]["pcm"] += 2.5
+        # clause-level evidence
+        C = sorted(set(cl))
+        cs = {c: dict.fromkeys(LANGS, 0.0) for c in C}
+        for i in range(n):
+            for l in LANGS:
+                cs[cl[i]][l] += sc[i][l]
+        # clause-final "o" in an ASCII clause = Pidgin particle (Yoruba "o" clauses carry other Yoruba evidence)
+        for i in range(n):
+            last = i == n - 1 or cl[i + 1] != cl[i]
+            if last and f[i] in PCM_AMBIG_END and asc[i]:
+                others = [j for j in range(n) if cl[j] == cl[i] and j != i]
+                if others and all(asc[j] for j in others) and cs[cl[i]]["yor"] < 2:
+                    sc[i]["pcm"] += 3; cs[cl[i]]["pcm"] += 3
+        if self.lid is not None:
+            for c in C:
+                ws = [toks[i][2] for i in range(n) if cl[i] == c]
+                for l, v in self._lid_scores(" ".join(ws)).items():
+                    cs[c][l] += 2.0 * v
+        out = []
+        for i in range(n):
+            c = cl[i]
+            ctx = _norm(cs[c])
+            nb = dict.fromkeys(LANGS, 0.0)        # neighbouring clauses, weaker
+            for d in (-1, 1):
+                if c + d in cs:
+                    for l, v in _norm(cs[c + d]).items():
+                        nb[l] += 0.3 * v
+            tot = {l: sc[i][l] + 1.5 * ctx[l] + nb[l] for l in LANGS}
+            # plain-ASCII words in a clause without Nigerian-language evidence: English or Pidgin
+            if asc[i] and max(sc[i][l] for l in NG) < 1.0:
+                ng_ctx = max(cs[c][l] for l in NG)
+                if ng_ctx < 1.5 * max(cs[c]["eng"], cs[c]["pcm"], 0.1):
+                    pc = cs[c]["pcm"] >= 2.5
+                    out.append("pcm" if pc else ("eng" if sc[i]["pcm"] < 4 else "pcm"))
+                    continue
+            best = max(tot, key=tot.get)
+            if tot[best] <= 0:
+                best = "eng"
+            out.append(best)
+        for i in range(n):                         # user overrides
+            if f[i] in self.lang_map:
+                out[i] = self.lang_map[f[i]]
+        return toks, out
+
+    def tag(self, text, force_lang=None):
+        text = unicodedata.normalize("NFC", text.strip())
+        if TAGGED.search(text):
+            return text                           # already tagged: leave alone
+        if force_lang:
+            return f"<{force_lang}>{text}</{force_lang}>"
+        toks, langs = self.word_langs(text)
+        if not toks:
+            return f"<eng>{text}</eng>"
+        # spans: consecutive same-language words; glue (spaces/punct) stays with the preceding word
+        spans, cur_l, a = [], langs[0], 0
+        for i in range(1, len(toks)):
+            if langs[i] != cur_l:
+                b = toks[i][0]
+                seg = text[a:b]
+                spans.append((cur_l, seg))
+                cur_l, a = langs[i], b
+        spans.append((cur_l, text[a:]))
+        lead = text[:toks[0][0]]
+        out = []
+        for k, (l, s) in enumerate(spans):
+            s2 = (lead + s) if k == 0 else s
+            core = s2.rstrip()
+            out.append(f"<{l}>{core}</{l}>")
+        return " ".join(out)
+
+
+def parse_map(s):
+    m = {}
+    for kv in (s or "").split(","):
+        if "=" in kv:
+            k, v = kv.split("=", 1)
+            v = v.strip()
+            if v not in LANGS:
+                raise ValueError(f"--lang_map: unknown language '{v}' (use {LANGS})")
+            m[k.strip()] = v
+    return m
+
+
+def spans_of(tagged):
+    return [(m.group(1), m.group(2)) for m in re.finditer(r"<(eng|ibo|hau|yor|pcm)>(.*?)</\1>", tagged, re.S)]
+
+
+def word_labels(tagged):
+    """gold/pred word-level labels from a tagged string."""
+    out = []
+    for l, s in spans_of(tagged):
+        out += [(fold(w), l) for _, _, w in tokenize(s)]
+    return out
+
+
+def strip_tags(tagged):
+    return re.sub(r"\s+", " ", re.sub(r"</?(eng|ibo|hau|yor|pcm)>", "", tagged)).strip()
+
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+    ap.add_argument("text")
+    ap.add_argument("--lang_map", default="", help='per-word overrides, e.g. "omo=pcm,sha=pcm"')
+    ap.add_argument("--force_lang", choices=LANGS)
+    ap.add_argument("--lid", choices=["none", "glotlid"], default="none")
+    a = ap.parse_args()
+    t = Tagger(None if a.lid == "none" else a.lid, parse_map(a.lang_map))
+    print(t.tag(a.text, a.force_lang))
+
+
+if __name__ == "__main__":
+    main()