Dynamic ASR v0.1.0: real (default) + codeswitch-experimental

LICENSE

@@ -0,0 +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.
+

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

@@ -0,0 +1,195 @@
+---
+license: cc-by-sa-4.0
+language:
+- en
+- pcm
+- ig
+- yo
+- ha
+pipeline_tag: automatic-speech-recognition
+library_name: fairseq2
+tags:
+- asr
+- speech-recognition
+- code-switching
+- nigerian-languages
+- igbo
+- yoruba
+- hausa
+- nigerian-pidgin
+- ctc
+- omnilingual-asr
+- streaming
+---
+
+# Dynamic ASR
+
+**Speech recognition for Nigerian English, Nigerian Pidgin, Igbo, Yoruba and Hausa — and for speech that switches between them inside a sentence.** By [Modular](https://huggingface.co/Modularcomputing), part of the Dynamic Accent project (see also the [Dynamic Accent TTS](https://huggingface.co/Modularcomputing/dynamic-accent-preview) and the [Dynamic Accent Duplex](https://huggingface.co/Modularcomputing/dynamic-accent-duplex) voice agent, which uses this model as its ASR).
+
+Dynamic ASR is Meta's [Omnilingual ASR](https://github.com/facebookresearch/omnilingual-asr) `omniASR_CTC_1B_v2` (1B-parameter wav2vec2 CTC, Apache-2.0) fully fine-tuned on **187.2 hours of licensed Nigerian speech**. The architecture, tokenizer and decoding are unchanged, so it is a drop-in replacement: language-agnostic (no language hint), greedy CTC, cheap enough to re-decode a growing buffer for streaming partials.
+
+**Status: experimental research release.** Read the limitations.
+
+## Two models
+
+| folder | what it is | use it when |
+|---|---|---|
+| `real/` (**default**) | fine-tuned on licensed real speech only | general use; as good as the experimental model on real code-switched speech and better on the selection dev set (mean CER 0.071 vs 0.074) |
+| `codeswitch-experimental/` | same recipe, plus ~4.99 h of **synthetic code-switched speech** made with our own Dynamic Accent Flash-DPO TTS (12 % of training audio) | research on synthetic code-switch data; see the ablation below before choosing it |
+
+Both are the same size (1B, ~1.95 GB bf16 state dict) and load the same way.
+
+## Results
+
+All test data is held out (never trained on). Scores are WER / CER after lower-casing and stripping punctuation; "no-diacritic" CER folds tone marks and dots. Decoding: greedy CTC with a Latin-script logit mask (the duplex setting).
+
+### Per language (duplex evaluation set: 229 clips — WAXAL test ibo/yor/hau/pcm × 40, FLEURS validation English × 29, 40 clips made by joining two held-out clips of different languages)
+
+**CER**
+
+| model | English | Pidgin | Igbo | Yoruba | Hausa | code-switch (concatenated) |
+|---|---|---|---|---|---|---|
+| omniASR_CTC_1B (stock; previous duplex ASR) | 0.065 | 0.169 | 0.206 | 0.191 | 0.090 | 0.128 |
+| omniASR_CTC_1B_v2 (stock base) | 0.024 | 0.123 | 0.133 | 0.135 | 0.106 | 0.104 |
+| **Dynamic ASR real** (`real/`) | 0.025 | 0.074 | 0.093 | 0.121 | 0.082 | 0.081 |
+| Dynamic ASR code-switch, experimental (`codeswitch-experimental/`) | 0.024 | 0.076 | 0.095 | 0.118 | 0.081 | 0.072 |
+
+**WER**
+
+| model | English | Pidgin | Igbo | Yoruba | Hausa | code-switch (concatenated) |
+|---|---|---|---|---|---|---|
+| omniASR_CTC_1B (stock; previous duplex ASR) | 0.144 | 0.380 | 0.636 | 0.498 | 0.299 | 0.330 |
+| omniASR_CTC_1B_v2 (stock base) | 0.097 | 0.336 | 0.456 | 0.401 | 0.317 | 0.292 |
+| **Dynamic ASR real** (`real/`) | 0.095 | 0.193 | 0.315 | 0.355 | 0.253 | 0.238 |
+| Dynamic ASR code-switch, experimental (`codeswitch-experimental/`) | 0.092 | 0.193 | 0.321 | 0.353 | 0.252 | 0.216 |
+
+### Real code-switched speech
+
+Licensed corpora contain little real code-switched speech, so we built the most honest test we could from held-out data: Igbo/Yoruba/Hausa utterances whose transcripts contain English words. *Natural conversational* = spontaneous Igbo speech with English (IgboSynCorp; those speakers were removed from training). *Insertional* = read Igbo/Yoruba/Hausa sentences with English names/terms (WAXAL test, NaijaS2ST dev). CER (no-diacritic CER in brackets), plus English-word recall (share of the reference's English words that appear in the output; strict exact-match, so low in absolute terms). The last column is a held-out set of synthetic code-switched clips (unseen texts) from the TTS.
+
+| model | real CS, all (n=94) | natural conversational Igbo-English (n=36) | insertional, read (n=58) | English-word recall | synthetic CS test (n=109) |
+|---|---|---|---|---|---|
+| omniASR_CTC_1B (stock; previous duplex ASR) | 0.224 (0.195) | 0.372 (0.326) | 0.178 (0.154) | 0.3009 | 0.159 (0.129) |
+| omniASR_CTC_1B_v2 (stock base) | 0.193 (0.172) | 0.353 (0.315) | 0.143 (0.128) | 0.2832 | 0.132 (0.118) |
+| **Dynamic ASR real** (`real/`) | 0.148 (0.137) | 0.248 (0.234) | 0.117 (0.107) | 0.3186 | 0.105 (0.091) |
+| Dynamic ASR code-switch, experimental (`codeswitch-experimental/`) | 0.149 (0.137) | 0.247 (0.234) | 0.119 (0.106) | 0.3805 | 0.038 (0.035) |
+
+### Did synthetic code-switched audio help?
+
+**Short answer: not on real code-switched speech — at least not measurably at this scale.** The two models share everything (data, seed, 4,000 steps) except that `codeswitch-experimental/` sampled 12 % of its audio from 2572 synthetic code-switched clips (4.99 h).
+
+- **Real code-switched speech (94 clips): a tie.** CER 0.148 (`real`) vs 0.149 (`codeswitch-experimental`); natural conversational Igbo–English 0.248 vs 0.247; insertional 0.117 vs 0.119; WER 0.436 vs 0.438. All differences are well inside the noise of a 94-clip test.
+- **English-word recall did go up**, 0.319 → 0.381 (share of the reference's English words the model writes out exactly). This is the one real-speech signal in favour of synthetic data: the model spells embedded English words as English more often. It did not translate into lower CER, and the test is small, so treat it as a lead, not a result.
+- **Per-language real tests: no change.** Differences of at most 0.003 CER on English, Pidgin, Igbo, Yoruba and Hausa (B slightly better on Yoruba/Hausa/English, slightly worse on Pidgin/Igbo). On the selection dev set (WAXAL validation + OpenSLR-70) the experimental model is slightly worse (mean CER 0.074 vs 0.071).
+- **Concatenated code-switch clips (real audio of two languages joined): small gain**, CER 0.081 → 0.072, WER 0.238 → 0.216. These are inter-sentential switches at a hard boundary, closer to what the synthetic data looks like than to natural speech.
+- **Synthetic code-switched test (109 held-out clips, unseen texts): large gain**, CER 0.105 → 0.038 (stock v2 base: 0.132). This mostly measures adaptation to our own TTS voices and recording style, so it is not evidence about real speakers.
+
+**Verdict:** synthetic code-switched audio from our TTS teaches the model the synthetic domain and slightly improves English-word spelling, but it does not lower error on real code-switched or monolingual speech. `real/` stays the default. A real natural code-switching test set (and more real code-switched training speech) is what would settle this.
+
+### AfriSwitch (in-the-wild code-switched speech) — pending
+
+We planned to score both models (and the stock v2 base) zero-shot on the Nigerian subsets of Intron Health's **AfriSwitch** benchmark (Igbo, Yoruba, Hausa, Nigerian Pidgin; evaluation only, never trained on). The dataset is gated with manual approval; we requested access on 2026-10-24 and it had not been granted when the models were finalised, so **no AfriSwitch scores are reported yet**. When we run it we will use our scorer's normalisation (lowercase, punctuation stripped; the paper scores diacritic-preserving WER, and our models drop Igbo tone marks, so Igbo WER will be inflated against its references and we will report the diacritic-folded figure alongside), corpus-level WER and CER.
+
+For reference, the paper's zero-shot WER (%) on these subsets (Table 2; systems given the matrix language, diacritics preserved):
+
+| language | Sahara V2.5 (best overall, avg 35.93 over 12 languages) | Omnilingual LLM 7B | Gemini 3.6 | ElevenLabs | Sahara V2 |
+|---|---|---|---|---|---|
+| Igbo | 46.94 | 75.52 | 75.88 | 65.37 | 83.58 |
+| Yoruba | 45.54 | 70.27 | 77.26 | 74.62 | 89.66 |
+| Hausa | 31.77 | 39.27 | 45.44 | 41.01 | 45.56 |
+| Nigerian Pidgin | 32.83 | 42.71 | 38.84 | 35.33 | 42.19 |
+
+AfriSwitch: Ashungafac, Awobade & Olatunji (Intron Health), *AfriSwitch: A Benchmark for In-the-Wild African Code-Switched Speech Recognition*, arXiv:2608.26434 (2026). Dataset: https://huggingface.co/datasets/intronhealth/AfriSwitch (dataset card licence: CC-BY-NC-SA-4.0; paper CC BY 4.0).
+
+```bibtex
+@misc{ashungafac2026afriswitch,
+  title={AfriSwitch: A Benchmark for In-the-Wild African Code-Switched Speech Recognition},
+  author={Ashungafac, Gabrial Zencha and Awobade, Busayo and Olatunji, Tobi},
+  year={2026}, eprint={2608.26434}, archivePrefix={arXiv}, primaryClass={cs.CL}
+}
+```
+
+## Usage
+
+```bash
+pip install "omnilingual-asr>=0.2.0" huggingface_hub soundfile scipy
+python transcribe.py audio.wav                      # real/ model (default)
+python transcribe.py audio.wav --model codeswitch   # codeswitch-experimental/
+```
+
+```python
+from transcribe import DynamicASR
+asr = DynamicASR("real")          # downloads real/model.pt; base architecture omniASR_CTC_1B_v2
+print(asr(waveform, sr))          # mono float32 array, any sample rate
+```
+
+The checkpoint is a plain PyTorch state dict (`{"model": state_dict, ...}`) for `omniASR_CTC_1B_v2`: `transcribe.py` loads the base through `fairseq2.models.hub.load_model("omniASR_CTC_1B_v2")` and overwrites its weights (the base download is needed once for the model skeleton and tokenizer). On `omnilingual-asr` 0.1.0, run `code/training/patch_v2.sh` (adjust the venv path) to register the v2 card first.
+
+**Streaming (duplex agent).** `code/asr_server.py` is the FastAPI worker used by Dynamic Accent Duplex: `POST /asr` with raw float32 16 kHz mono → `{"text", "ms"}`; partial hypotheses come from re-decoding the growing buffer every 0.64 s (~30–90 ms per call on an H100).
+
+```bash
+cd code && ASR_MODEL=dynamic-asr uvicorn asr_server:app --port 8002      # real/ (default)
+#          ASR_MODEL=dynamic-asr-cs  -> codeswitch-experimental/
+#          ASR_MODEL=omni            -> stock omniASR_CTC_1B (also the automatic fallback if Dynamic ASR cannot load)
+```
+
+**Output format:** lower-case, no punctuation, digits as spoken or written in the training transcripts. Igbo is written **without tone marks** (dotted letters ị ọ ụ ṅ kept, standard everyday orthography); Yoruba **with** tone marks; Hausa as in the sources (hooked letters often missing, as in WAXAL).
+
+## Training data (licensed only)
+
+| source | language | hours used | licence |
+|---|---|---|---|
+| bible_hau | hau | 27.01 | CC-BY-SA-4.0 (Open.Bible / Biblica) |
+| bible_yor | yor | 20.43 | CC-BY-SA-4.0 (Open.Bible / Biblica) |
+| fleurs | eng | 5.63 | CC-BY-4.0 (Google FLEURS) |
+| fleurs | hau | 7.6 | CC-BY-4.0 (Google FLEURS) |
+| fleurs | ibo | 6.09 | CC-BY-4.0 (Google FLEURS) |
+| fleurs | yor | 5.74 | CC-BY-4.0 (Google FLEURS) |
+| igboapi | ibo | 34.98 | CC-BY-4.0 (IgboAPI) |
+| igbosyncorp | ibo | 2.16 | CC0-1.0 (IgboSynCorp) |
+| naijas2st_igbo | ibo | 14.81 | CC-BY-4.0 (NaijaS2ST) |
+| open_bible_igbo | ibo | 22.02 | CC-BY-SA-4.0 (Open.Bible / Biblica) |
+| slr70 | eng | 4.0 | CC-BY-SA-4.0 (OpenSLR 70, Nigerian English) |
+| slr86 | yor | 2.98 | CC-BY-SA-4.0 (OpenSLR 86, Yoruba) |
+| waxal | hau | 6.58 | CC-BY-SA-4.0 (WAXAL Media Trust subsets) |
+| waxal | ibo | 10.58 | CC-BY-SA-4.0 (WAXAL Media Trust subsets) |
+| waxal | pcm | 6.07 | CC-BY-SA-4.0 (WAXAL Media Trust subsets) |
+| waxal | yor | 10.48 | CC-BY-SA-4.0 (WAXAL Media Trust subsets) |
+
+Per-epoch sampling down-weights Bible narration and dictionary reads and up-weights Pidgin (≈ 115.6 h per epoch; Igbo 39, Yoruba 26, Hausa 24, Pidgin 18, English 8 h). Held out and never trained on: WAXAL test + validation, FLEURS dev/test, NaijaS2ST dev, and every IgboSynCorp speaker whose speech contained English.
+
+**Synthetic code-switched data (`codeswitch-experimental/` only).** 2797 code-switched sentences (English with Igbo/Yoruba/Hausa/Pidgin, Pidgin with Igbo/Yoruba/Hausa; spans tagged per language) written by Qwen3-30B-A3B-Instruct-2507 (Apache-2.0), read by the Dynamic Accent **Flash-DPO** TTS (itself trained on licensed data only) in 32 voices (zero-shot prompts from WAXAL speakers), with random speaking rate (slow/normal/fast), speed perturbation 0.9–1.1, and on-the-fly reverb, phone/cheap-mic band-limiting, clipping and noise (SNR 5–30 dB). Clips the stock model could not read (CER > 0.35, i.e. TTS failures) were dropped: 2572 clips / 4.99 h kept for training, 109 unseen-text clips kept as the synthetic test.
+
+**Not used, on licence grounds:** African Voices / WaZoBia, NaijaVoices (non-commercial), AfriSpeech-200 (non-commercial). Common Voice was not used (not fetched). No African Voices–trained TTS was used to make synthetic data.
+
+## Training
+
+Full fine-tune of all transformer layers (CNN feature encoder frozen), bf16 autocast, AdamW (lr 2e-5, β 0.9/0.98, wd 0.01), 300 warm-up steps then linear decay, 4000 steps × 160 s of audio per batch (~177.8 h of audio seen), CTC loss, LayerDrop 0.1, 15 % light augmentation of real clips. One H100; the two models differ only in the synthetic share (0 % vs 12 %), same seed and steps. Checkpoint selected on a WAXAL-validation/OpenSLR-70 dev set. Training, data-prep, synthesis and evaluation code is in `code/training/`.
+
+## Limitations
+
+- **Small, read-speech-heavy test sets.** 40 clips per language; real code-switched test = 94 clips. Differences of ~0.01 CER are within noise.
+- **Convention matching is part of the gain.** WAXAL train and test share transcription conventions (e.g. Igbo without tones, Hausa without hooked letters, digits as numerals); some of the improvement on WAXAL tests is learning those conventions, not only better hearing.
+- **Little natural conversational speech** (≈2 h IgboSynCorp in training; none for Yoruba/Hausa). Expect higher error rates on spontaneous, noisy, phone or far-field speech, and on accents unlike the training speakers.
+- **Code-switching is the weakest area.** Real intra-sentential code-switched test data is scarce; the "natural conversational" CER is still ~0.25. English inside Igbo/Yoruba/Hausa is often spelt phonetically.
+- Yoruba tone marks are frequently wrong or missing (Yoruba CER without diacritics is much lower than with); Igbo output has no tone marks by design.
+- Not evaluated yet on the developer's own voice or other unseen real users; not evaluated for fairness across gender/age/region.
+- No punctuation or casing (unlike the stock v2 model's written output).
+- Not for high-stakes use (medical, legal, emergency) without human review.
+
+## Licence and attribution
+
+- **Weights: CC-BY-SA-4.0.** Several training sources are CC-BY-SA-4.0 (WAXAL Media Trust subsets, Open.Bible, OpenSLR 70/86), so the fine-tuned weights are shared alike under CC-BY-SA-4.0. The base model omniASR_CTC_1B_v2 is Apache-2.0 (© Meta Platforms).
+- **Code (`transcribe.py`, `code/`): Apache-2.0** (`LICENSE-CODE`).
+- Attribution: Google WAXAL (Media Trust subsets), Google FLEURS (CC-BY-4.0), OpenSLR 70 Nigerian English and OpenSLR 86 Yoruba (Google, CC-BY-SA-4.0), Open.Bible / Biblica (CC-BY-SA-4.0), NaijaS2ST (CC-BY-4.0), IgboAPI (CC-BY-4.0), IgboSynCorp (CC0-1.0, Harvard Dataverse doi:10.7910/DVN/YB9FWK), Meta Omnilingual ASR (Apache-2.0), Qwen3 (Apache-2.0).
+
+## Citation
+
+```bibtex
+@misc{modular2026dynamicasr,
+  title  = {Dynamic ASR: Nigerian multilingual and code-switched speech recognition},
+  author = {Modular (Chima Emmanuel Akuegbo)},
+  year   = {2026},
+  url    = {https://huggingface.co/Modularcomputing/dynamic-asr}
+}
+```

code/asr_server.py

@@ -0,0 +1,95 @@
+"""Streaming-capable ASR worker (runs in ~/venv_omni): Meta Omnilingual ASR CTC (Apache-2.0), language-agnostic decoding,
+so code-switched speech needs no language hint. POST /asr  body = raw float32 mono 16 kHz  -> {"text", "ms"}.
+Partial hypotheses = re-decoding the growing buffer (CTC is cheap: ~10-40 ms per call on H100)."""
+import os, time, numpy as np, torch
+import soundfile  # noqa: F401  (loads the bundled libsndfile before fairseq2n looks for it)
+from fastapi import FastAPI, Request
+from omnilingual_asr.models.inference.pipeline import ASRInferencePipeline
+# ASR model selection (env):
+#   ASR_MODEL=dynamic-asr (default) -> Dynamic ASR real-speech model  (Modularcomputing/dynamic-asr, real/model.pt)
+#   ASR_MODEL=dynamic-asr-cs        -> Dynamic ASR experimental code-switch model (codeswitch-experimental/model.pt)
+#   ASR_MODEL=omni                  -> stock Meta omniASR (ASR_CARD, default omniASR_CTC_1B)
+#   ASR_CKPT=<local .pt>            -> any local fine-tuned state_dict on top of ASR_CARD
+# Dynamic ASR = fine-tuned omniASR_CTC_1B_v2 (same CTC architecture; needs omnilingual-asr>=0.2.0, or patch_v2.sh on 0.1.0).
+# If the Dynamic ASR weights cannot be loaded, the worker falls back to stock omniASR_CTC_1B.
+LC = {"ibo": "ibo_Latn", "yor": "yor_Latn", "hau": "hau_Latn", "pcm": "pcm_Latn", "eng": "eng_Latn"}
+import unicodedata, types
+import omnilingual_asr.models.inference.pipeline as PL
+DYN = {"dynamic-asr": "real/model.pt", "dynamic-asr-cs": "codeswitch-experimental/model.pt"}
+MODEL = os.environ.get("ASR_MODEL", "dynamic-asr")
+CKPT = os.environ.get("ASR_CKPT", "")
+CARD = os.environ.get("ASR_CARD", "omniASR_CTC_1B_v2" if (MODEL in DYN or CKPT) else "omniASR_CTC_1B")
+
+def _load():
+    global CARD, CKPT, MODEL
+    if MODEL in DYN:
+        CARD = "omniASR_CTC_1B_v2"          # Dynamic ASR is always on the v2 base
+    if MODEL in DYN and not CKPT:
+        try:
+            from huggingface_hub import hf_hub_download
+            CKPT = hf_hub_download(os.environ.get("ASR_REPO", "Modularcomputing/dynamic-asr"), DYN[MODEL])
+        except Exception as e:
+            print("Dynamic ASR download failed, falling back to omniASR_CTC_1B:", repr(e)[:200], flush=True)
+            CARD, MODEL, CKPT = "omniASR_CTC_1B", "omni", ""
+    try:
+        pipe = ASRInferencePipeline(model_card=CARD)
+        if CKPT:
+            ck = torch.load(CKPT, map_location="cpu", weights_only=False)
+            print("ASR_CKPT", CKPT, pipe.model.load_state_dict({k: v.to(pipe.dtype) for k, v in ck["model"].items()}), flush=True)
+            del ck
+        return pipe
+    except Exception as e:
+        if CARD == "omniASR_CTC_1B":
+            raise
+        print("Dynamic ASR load failed (", repr(e)[:200], ") -> falling back to omniASR_CTC_1B", flush=True)
+        CARD, MODEL, CKPT = "omniASR_CTC_1B", "omni", ""
+        return ASRInferencePipeline(model_card=CARD)
+
+p = _load()
+LATIN_ONLY = os.environ.get("ASR_LATIN_ONLY", "1") == "1"
+
+def _ok(sx):
+    for c in sx:
+        cat = unicodedata.category(c)
+        if cat[0] == "M":                                   # combining marks: only the generic combining diacritics block
+            if not ("\u0300" <= c <= "\u036f"):
+                return False
+        elif c.isalpha() and "LATIN" not in unicodedata.name(c, ""):
+            return False
+    return True
+
+_BAD = {}
+def _apply_latin(self, batch):
+    """CTC greedy decoding restricted to Latin-script tokens: all five target languages are written in Latin script, and the
+    language-agnostic CTC otherwise drifts into Arabic/other scripts on accented or synthetic speech."""
+    bl = PL.BatchLayout(batch.source_seqs.shape, seq_lens=batch.source_seq_lens, device=batch.source_seqs.device)
+    logits, bl_out = self.model(batch.source_seqs, bl)
+    V = logits.shape[-1]
+    if V not in _BAD:
+        bad = [i for i in range(V) if not _ok(self.token_decoder(torch.tensor([i])))]
+        _BAD[V] = torch.tensor(bad, dtype=torch.long, device=logits.device); print("latin mask: blocked", len(bad), "of", V, flush=True)
+    logits[..., _BAD[V]] = -float("inf")
+    pred = torch.argmax(logits, dim=-1); out = []
+    for i in range(pred.shape[0]):
+        seq = pred[i][: bl_out.seq_lens[i]]; m = torch.ones(seq.shape[0], dtype=torch.bool, device=seq.device); m[1:] = seq[1:] != seq[:-1]
+        out.append(self.token_decoder(seq[m]))
+    return out
+
+if LATIN_ONLY and "CTC" in CARD:
+    p._apply_model_wav2vec2asr = types.MethodType(_apply_latin, p)
+app = FastAPI()
+w = np.zeros(16000, np.float32); p.transcribe([{"waveform": w, "sample_rate": 16000}], batch_size=1)
+
+@app.post("/asr")
+async def asr(req: Request, lang: str = ""):
+    y = np.frombuffer(await req.body(), dtype=np.float32).copy()
+    if len(y) < 1600:
+        return {"text": "", "ms": 0.0}
+    t = time.perf_counter()
+    kw = {} if "CTC" in CARD or not lang else dict(lang=[LC.get(lang, "eng_Latn")])
+    h = p.transcribe([{"waveform": y, "sample_rate": 16000}], batch_size=1, **kw)[0]
+    return {"text": h, "ms": round(1000 * (time.perf_counter() - t), 1), "card": CARD}
+
+@app.get("/health")
+def health():
+    return {"ok": True, "model": MODEL, "card": CARD, "ckpt": CKPT}

code/training/build_synth.py

@@ -0,0 +1,23 @@
+"""Merge synth parts, split held-out synthetic test (every 25th text id), and (after eval_asr.py --hyp_only) filter TTS failures.
+python3 build_synth.py merge   -> /scratch/asr/synth/all.jsonl
+python3 build_synth.py filter  -> /scratch/asr/synth/{manifest.jsonl (train), test.jsonl}  (keep clips the stock v2 model reads at CER <= 0.35)"""
+import json, glob, sys, collections
+D = "/scratch/asr/synth"
+if sys.argv[1] == "merge":
+    R = [json.loads(l) for f in sorted(glob.glob(f"{D}/parts/part_*.jsonl")) for l in open(f)]
+    R = [r for r in R if 1.0 <= r["dur"] <= 20 and r["text"]]
+    with open(f"{D}/all.jsonl", "w") as f:
+        for r in R: f.write(json.dumps(r, ensure_ascii=False) + "\n")
+    print("merged", len(R), "hours", round(sum(r["dur"] for r in R) / 3600, 2))
+else:
+    R = [json.loads(l) for l in open(f"{D}/all.jsonl")]; H = {json.loads(l)["id"]: json.loads(l) for l in open(f"{D}/all.jsonl.hyp.jsonl")}
+    tr, te, drop = [], [], collections.Counter()
+    for r in R:
+        c = H[r["id"]]["cer"]; r["base_cer"] = c
+        if c > 0.35: drop[r["pair"]] += 1; continue
+        (te if int(r["id"][2:]) % 25 == 0 else tr).append(r)
+    for name, X in (("manifest", tr), ("test", te)):
+        with open(f"{D}/{name}.jsonl", "w") as f:
+            for r in X: f.write(json.dumps(r, ensure_ascii=False) + "\n")
+    print("train", len(tr), round(sum(r["dur"] for r in tr) / 3600, 2), "h; test", len(te), "; dropped (base CER>0.35)", sum(drop.values()), dict(drop))
+    print("pairs train", collections.Counter(r["pair"] for r in tr))

code/training/eval_asr.py

@@ -0,0 +1,100 @@
+"""Evaluate an omniASR CTC model (+ optional fine-tuned checkpoint) with the duplex decoding (greedy CTC + Latin-script mask).
+~/venv_omni/bin/python eval_asr.py --tag NAME [--card omniASR_CTC_1B_v2] [--ckpt /scratch/asr/runs/A/best.pt] [--sets duplex,realcs,synthtest,ace]
+  --hyp_only MANIFEST : just transcribe a manifest (used for the synthetic-data CER filter) -> MANIFEST.hyp.jsonl
+Writes /scratch/asr/eval/<tag>.json (scores) and <tag>_hyps.jsonl."""
+import argparse, json, os, re, unicodedata, time, numpy as np, torch, torch.nn.functional as F
+import soundfile as sf
+from scipy.signal import resample_poly
+from fairseq2.models.hub import load_model
+from fairseq2.data.tokenizers.hub import load_tokenizer
+from fairseq2.nn.batch_layout import BatchLayout
+import jiwer
+ap = argparse.ArgumentParser()
+ap.add_argument("--tag", default="x"); ap.add_argument("--card", default="omniASR_CTC_1B_v2"); ap.add_argument("--ckpt", default="")
+ap.add_argument("--sets", default="duplex,realcs,synthtest,ace"); ap.add_argument("--hyp_only", default=""); ap.add_argument("--latin", type=int, default=1)
+a = ap.parse_args()
+dev = torch.device("cuda")
+model = load_model(a.card, device=dev, dtype=torch.bfloat16).eval()
+if a.ckpt:
+    ck = torch.load(a.ckpt, map_location="cuda", weights_only=False); print("load", a.ckpt, model.load_state_dict(ck["model"]), ck.get("step"), ck.get("dev"))
+tok = load_tokenizer(a.card); dec = tok.create_decoder(skip_special_tokens=True)
+V = model.final_proj.weight.shape[0]
+def _ok(sx):
+    for c in sx:
+        cat = unicodedata.category(c)
+        if cat[0] == "M":
+            if not ("\u0300" <= c <= "\u036f"): return False
+        elif c.isalpha() and "LATIN" not in unicodedata.name(c, ""): return False
+    return True
+BAD = torch.tensor([i for i in range(V) if not _ok(str(dec(torch.tensor([i]))))], dtype=torch.long, device=dev) if a.latin else None
+print("latin mask blocks", 0 if BAD is None else len(BAD), "of", V)
+
+def load(r):
+    if r.get("wav"):
+        y, sr = sf.read(r["wav"], dtype="float32")
+        if y.ndim > 1: y = y.mean(1)
+        return y if sr == 16000 else resample_poly(y, 16000, sr).astype(np.float32)
+    m = np.memmap(r["pcm"], dtype=np.int16, mode="r"); y = m[r["off24"]: r["off24"] + r["n24"]].astype(np.float32) / 32768
+    return resample_poly(y, 2, 3).astype(np.float32)
+
+@torch.no_grad()
+def transcribe(items, bs=8):
+    out, lat = [], []
+    order = sorted(range(len(items)), key=lambda i: items[i].get("dur", 0)); res = [None] * len(items)
+    for s in range(0, len(order), bs):
+        idx = order[s: s + bs]; ys = [torch.from_numpy(load(items[i])) for i in idx]
+        ys = [F.layer_norm(y, y.shape) for y in ys]; L = [len(y) for y in ys]
+        X = torch.zeros(len(ys), max(L)); [X[k, : len(y)].copy_(y) for k, y in enumerate(ys)]
+        X = X.to(dev, torch.bfloat16); bl = BatchLayout(X.shape, seq_lens=L, device=dev)
+        lg, blo = model(X, bl)
+        if BAD is not None: lg[..., BAD] = -float("inf")
+        pr = lg.argmax(-1)
+        for k, i in enumerate(idx):
+            sq = pr[k, : blo.seq_lens[k]]; m = torch.ones_like(sq, dtype=torch.bool); m[1:] = sq[1:] != sq[:-1]
+            res[i] = str(dec(sq[m]))
+    return res
+
+def n(s):
+    s = unicodedata.normalize("NFC", s.lower()); s = re.sub(r"[^\w\s']", " ", s); return re.sub(r"\s+", " ", s).strip()
+def fold(s): return "".join(c for c in unicodedata.normalize("NFD", s) if unicodedata.category(c) != "Mn")
+def score(refs, hyps):
+    R = [n(r) or "x" for r in refs]; H = [n(h) or "x" for h in hyps]
+    return dict(n=len(R), wer=round(jiwer.wer(R, H), 4), cer=round(jiwer.cer(R, H), 4),
+                cer_nodiac=round(jiwer.cer([fold(r) for r in R], [fold(h) for h in H]), 4))
+
+if a.hyp_only:
+    items = [json.loads(l) for l in open(a.hyp_only)]; hy = transcribe(items, 16)
+    with open(a.hyp_only + ".hyp.jsonl", "w") as f:
+        for r, h in zip(items, hy):
+            f.write(json.dumps(dict(id=r["id"], hyp=h, cer=round(jiwer.cer(n(r["text"]) or "x", n(h) or "x"), 4)), ensure_ascii=False) + "\n")
+    print("HYP_DONE", len(items)); raise SystemExit
+
+os.makedirs("/scratch/asr/eval", exist_ok=True); res = {}; allh = []
+sets = a.sets.split(",")
+if "duplex" in sets:   # the duplex REPORT set: WAXAL test x40 ibo/yor/hau/pcm, FLEURS val eng x29, 40 concatenated cs clips
+    R = [json.loads(l) for l in open("/scratch/duplex/asr_eval/list.jsonl")]; H = transcribe(R)
+    for lg in ["eng", "pcm", "ibo", "yor", "hau", "cs"]:
+        ix = [i for i, r in enumerate(R) if r["lang"] == lg]; res[f"duplex_{lg}"] = score([R[i]["text"] for i in ix], [H[i] for i in ix])
+    allh += [dict(set="duplex", id=r["id"], ref=r["text"], hyp=h) for r, h in zip(R, H)]
+if "realcs" in sets:   # real speech with English inside Igbo/Yoruba/Hausa sentences (held out)
+    R = [json.loads(l) for l in open("/scratch/asr/data/realcs_test.jsonl")]; H = transcribe(R)
+    res["realcs_all"] = score([r["text"] for r in R], H)
+    for k in ["natural_conversational", "insertional"]:
+        ix = [i for i, r in enumerate(R) if r["cs_kind"] == k]; res[f"realcs_{k}"] = score([R[i]["text"] for i in ix], [H[i] for i in ix])
+    hit = tot = 0   # English-word recall: share of the English words in the reference that appear in the hypothesis
+    for r, h in zip(R, H):
+        hw = n(h).split()
+        for w in r["cs_eng"]:
+            tot += 1; hit += w in hw
+    res["realcs_eng_word_recall"] = round(hit / max(1, tot), 4)
+    allh += [dict(set="realcs", id=r["id"], ref=r["text"], hyp=h) for r, h in zip(R, H)]
+if "synthtest" in sets and os.path.exists("/scratch/asr/synth/test.jsonl"):   # held-out synthetic cs texts (TTS domain)
+    R = [json.loads(l) for l in open("/scratch/asr/synth/test.jsonl")]; H = transcribe(R); res["synthtest"] = score([r["text"] for r in R], H)
+    allh += [dict(set="synthtest", id=r["id"], ref=r["text"], hyp=h) for r, h in zip(R, H)]
+if "ace" in sets and os.path.exists("/scratch/asr/ace/list.jsonl"):   # ACE's own recordings, when available
+    R = [json.loads(l) for l in open("/scratch/asr/ace/list.jsonl")]; H = transcribe(R, 1); res["ace"] = score([r["text"] for r in R], H)
+    allh += [dict(set="ace", id=r["id"], ref=r["text"], hyp=h) for r, h in zip(R, H)]
+json.dump(dict(tag=a.tag, card=a.card, ckpt=a.ckpt, scores=res), open(f"/scratch/asr/eval/{a.tag}.json", "w"), indent=1)
+with open(f"/scratch/asr/eval/{a.tag}_hyps.jsonl", "w") as f:
+    for h in allh: f.write(json.dumps(h, ensure_ascii=False) + "\n")
+print(json.dumps(res, indent=1)); print("EVAL_DONE", a.tag)

code/training/gen_cs_text.py

@@ -0,0 +1,61 @@
+"""Generate intra-sentence code-switched Nigerian sentences (tagged spans) with Qwen3-30B-A3B-Instruct-2507-FP8 (Apache-2.0), offline vLLM.
+-> /scratch/asr/synth/cs_text.jsonl  {id, tagged, pair, topic}
+/scratch/vllm/bin/python gen_cs_text.py [n_prompts_per_pair]"""
+import json, os, random, re, sys
+from vllm import LLM, SamplingParams
+random.seed(5)
+OUT = "/scratch/asr/synth"; os.makedirs(OUT, exist_ok=True)
+NAME = {"eng": "English", "ibo": "Igbo", "yor": "Yoruba", "hau": "Hausa", "pcm": "Nigerian Pidgin"}
+# (matrix, embedded): matrix language carries most of the sentence, the other is switched in mid-sentence
+PAIRS = [("ibo", "eng"), ("eng", "ibo"), ("yor", "eng"), ("eng", "yor"), ("hau", "eng"), ("eng", "hau"),
+         ("pcm", "ibo"), ("pcm", "yor"), ("pcm", "hau"), ("eng", "pcm"), ("pcm", "eng")]
+TOPICS = ["greeting a friend and asking about family", "a phone call about being late", "buying food at the market and bargaining",
+          "taking a bus, keke or okada in traffic", "school, exams and lecturers", "work, a boss and a deadline", "church or mosque on the weekend",
+          "football, a Premier League match", "a bank transfer that failed", "fixing a phone or laptop", "a wedding or naming ceremony",
+          "cooking jollof rice or soup at home", "rain, heat and NEPA light going off", "a hospital visit", "a WhatsApp group chat argument",
+          "a job interview", "rent and the landlord", "music, Afrobeats and a concert", "a birthday party plan", "giving directions in town",
+          "asking an AI assistant for help", "a business deal with a customer", "complaining about fuel prices", "studying for JAMB or WAEC",
+          "a village visit for Christmas", "a doctor's advice", "ordering food delivery", "a startup and a product launch"]
+STY = ["casual and friendly", "excited", "polite and respectful", "complaining", "joking", "in a hurry", "telling a story"]
+
+def prompt(m, e, topic, sty, n=15):
+    sw = (f"Most of each sentence is in {NAME[m]}, with {NAME[e]} words or phrases switched in mid-sentence, the way Nigerians really talk."
+          if m != "eng" else f"Most of each sentence is in English, with {NAME[e]} words, phrases or clauses switched in mid-sentence, the way Nigerians really talk.")
+    return (f"Write {n} different spoken sentences a Nigerian might say about: {topic}. Tone: {sty}. {sw}\n"
+            f"Rules: every sentence switches language at least once INSIDE the sentence (not just at the start or end); 6 to 24 words; "
+            f"natural everyday speech; use correct standard spelling ({'Yoruba with tone marks and dots' if 'yor' in (m, e) else ''}"
+            f"{'Igbo with dotted vowels (ị ọ ụ ṅ)' if 'ibo' in (m, e) else ''}{'Hausa with hooked letters (ɓ ɗ ƙ ƴ) where needed' if 'hau' in (m, e) else ''}); "
+            f"write numbers as words; no emojis, no hashtags, no quotation marks.\n"
+            f"Wrap every span in its language tag: <{m}>...</{m}> and <{e}>...</{e}>. Example format: "
+            f"<{m}>...</{m}> <{e}>...</{e}> <{m}>...</{m}>\n"
+            f"Output ONLY a JSON array of {n} strings.")
+
+N = int(sys.argv[1]) if len(sys.argv) > 1 else 45
+jobs = []
+for m, e in PAIRS:
+    k = N if "pcm" not in (m, e) or (m, e) in (("pcm", "eng"), ("eng", "pcm")) else int(N * 0.6)
+    for i in range(k):
+        t, s = random.choice(TOPICS), random.choice(STY); jobs.append(dict(pair=f"{m}+{e}", topic=t, msg=prompt(m, e, t, s)))
+llm = LLM("Qwen/Qwen3-30B-A3B-Instruct-2507-FP8", gpu_memory_utilization=float(os.environ.get("GPU_UTIL", "0.42")), max_model_len=4096,
+          enforce_eager=False)
+sp = SamplingParams(temperature=0.9, top_p=0.95, max_tokens=1800, seed=0)
+outs = llm.chat([[{"role": "user", "content": j["msg"]}] for j in jobs], sp)
+TAG = re.compile(r"<(eng|ibo|yor|hau|pcm)>(.*?)</\1>", re.S)
+seen, n_ok, n_bad = set(), 0, 0
+with open(f"{OUT}/cs_text.jsonl", "w") as f:
+    for j, o in zip(jobs, outs):
+        txt = o.outputs[0].text; i0, i1 = txt.find("["), txt.rfind("]")
+        try: arr = json.loads(txt[i0: i1 + 1])
+        except Exception: n_bad += 1; continue
+        for s in arr:
+            if not isinstance(s, str): continue
+            spans = [(l, re.sub(r"\s+", " ", t).strip()) for l, t in TAG.findall(s)]
+            spans = [(l, t) for l, t in spans if t]
+            langs = [l for l, _ in spans]; nw = sum(len(t.split()) for _, t in spans)
+            if len(spans) < 2 or len(set(langs)) < 2 or not (5 <= nw <= 30) or set(langs) - set(j["pair"].split("+")): n_bad += 1; continue
+            if any(len(t) > 200 for _, t in spans) or re.search(r"[<>\"#@]", "".join(t for _, t in spans)): n_bad += 1; continue
+            key = " ".join(t for _, t in spans).lower()
+            if key in seen: continue
+            seen.add(key); tagged = " ".join(f"<{l}>{t}</{l}>" for l, t in spans)
+            f.write(json.dumps(dict(id=f"cs{n_ok:05d}", tagged=tagged, pair=j["pair"], topic=j["topic"]), ensure_ascii=False) + "\n"); n_ok += 1
+print("CS_TEXT_DONE ok", n_ok, "bad", n_bad, "prompts", len(jobs))

code/training/patch_v2.sh

@@ -0,0 +1,39 @@
+#!/bin/bash
+# register omniASR_CTC_1B_v2 (arch 1b_v2 = 1b with vocab 10288, tokenizer omniASR_tokenizer_written_v2) in venv_omni 0.1.0
+P=/home/dev/venv_omni/lib/python3.12/site-packages/omnilingual_asr
+grep -q '"1b_v2"' $P/models/wav2vec2_asr/config.py || python3 - <<'PY'
+p="/home/dev/venv_omni/lib/python3.12/site-packages/omnilingual_asr/models/wav2vec2_asr/config.py"
+s=open(p).read()
+add='''
+    @arch("1b_v2", advanced=True)
+    def _1b_asr_v2(resolver: DependencyResolver) -> Wav2Vec2AsrConfig:
+        config = get_config(resolver, Wav2Vec2AsrConfig, "base_10h")
+        config.encoder_config = get_config(resolver, Wav2Vec2Config, "1b").encoder_config
+        config.encoder_config.dropout_p = 0.0
+        config.encoder_config.attn_dropout_p = 0.0
+        config.encoder_config.ffn_inner_dropout_p = 0.1
+        config.encoder_config.layer_drop_p = 0.1
+        config.use_masking = False
+        config.max_temporal_mask_prob = 0.0
+        config.max_spatial_mask_prob = 0.0
+        config.target_vocab_size = 10288
+        return config
+'''
+i=s.index('    @arch("1b", advanced=True)')
+s=s[:i]+add.lstrip("\n")+"\n"+s[i:]
+open(p,"w").write(s); print("arch patched")
+PY
+cat > $P/cards/models/rc_models_v2_local.yaml <<'Y'
+name: omniASR_tokenizer_written_v2
+tokenizer_family: char_tokenizer
+tokenizer: https://dl.fbaipublicfiles.com/mms/omniASR_tokenizer_written_v2.model
+
+---
+
+name: omniASR_CTC_1B_v2
+model_family: wav2vec2_asr
+model_arch: 1b_v2
+checkpoint: https://dl.fbaipublicfiles.com/mms/omniASR-CTC-1B-v2.pt
+tokenizer_ref: omniASR_tokenizer_written_v2
+Y
+echo card written

code/training/pipeline2.sh

@@ -0,0 +1,22 @@
+#!/bin/bash
+# After A_real finishes: baseline/A evals, finish synthesis (deadline), filter, train B_synth, final evals.
+. ~/asr/env.sh; cd ~/asr; L=/scratch/logs; E="PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True"
+while pgrep -f "runs/A_real " >/dev/null || pgrep -f "runs/A_real$" >/dev/null || pgrep -f "train_asr.py --run /scratch/asr/runs/A_real" >/dev/null; do sleep 30; done
+echo "A finished $(date -u)"; grep -E "TRAIN_DONE|Error" $L/train_A.log | tail -2
+nohup ~/venv/bin/python -u synth_cs.py 4 5 > $L/synth_4.log 2>&1 &
+~/venv_omni/bin/python -u eval_asr.py --tag v1_stock --card omniASR_CTC_1B --sets duplex,realcs > $L/eval_v1.log 2>&1; echo "eval v1 $?"
+~/venv_omni/bin/python -u eval_asr.py --tag v2_stock --card omniASR_CTC_1B_v2 --sets duplex,realcs > $L/eval_v2.log 2>&1; echo "eval v2 $?"
+~/venv_omni/bin/python -u eval_asr.py --tag A_real --ckpt /scratch/asr/runs/A_real/best.pt --sets duplex,realcs > $L/eval_A.log 2>&1; echo "eval A $?"
+DL=$(date -u -d "2026-10-24 18:40" +%s)
+while [ $(date -u +%s) -lt $DL ]; do
+  n=$(grep -l SYNTH_DONE $L/synth_[0-4].log 2>/dev/null | wc -l); [ "$n" -ge 5 ] && break; sleep 30; done
+pkill -f synth_cs.py; sleep 5; echo "synth stopped $(date -u)"
+python3 build_synth.py merge
+~/venv_omni/bin/python -u eval_asr.py --hyp_only /scratch/asr/synth/all.jsonl > $L/synth_hyp.log 2>&1; echo "hyp $?"
+python3 build_synth.py filter
+env PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True MEM_FRAC=0.9 ~/venv_omni/bin/python -u train_asr.py --run /scratch/asr/runs/B_synth --steps 4000 --batch_sec 160 --max_sec 18 --val_every 500 \
+  --synth /scratch/asr/synth/manifest.jsonl --synth_frac 0.12 > $L/train_B.log 2>&1; echo "train B $?"
+~/venv_omni/bin/python -u eval_asr.py --tag B_synth --ckpt /scratch/asr/runs/B_synth/best.pt > $L/eval_B.log 2>&1; echo "eval B $?"
+for t in "v1_stock --card omniASR_CTC_1B" "v2_stock --card omniASR_CTC_1B_v2" "A_real --ckpt /scratch/asr/runs/A_real/best.pt"; do
+  set -- $t; ~/venv_omni/bin/python -u eval_asr.py --tag ${1}_st $2 $3 --sets synthtest,ace > $L/eval_st_$1.log 2>&1; echo "eval st $1 $?"; done
+echo PIPE_DONE $(date -u)

code/training/prep_data.py

@@ -0,0 +1,112 @@
+"""Build ASR fine-tuning manifests from licensed, public-safe Nigerian speech already on dynacc-h100.
+-> /scratch/asr/data/{train.jsonl, dev.jsonl, realcs_test.jsonl, stats.json}
+Each row: id, src, lang, spk, text (ASR target, normalized), dur, pcm (24 kHz int16 file), off24, n24, licence.
+Excluded (never touched): African Voices / WaZoBia, NaijaVoices, AfriSpeech (NC), Flash-AV. Held out: WAXAL test+validation,
+FLEURS dev/test/val, NaijaS2ST dev, IgboSynCorp speakers that produced natural code-switched clips (-> realcs test).
+Target normalization (matches duplex asr_score.norm): NFC, lowercase, punctuation -> space (apostrophes kept);
+Igbo: tone marks (acute/grave/macron) stripped, dot-below and n-dot kept (WAXAL/standard Igbo orthography);
+Yoruba/Hausa: kept as written by the source."""
+import json, os, re, unicodedata, random, collections
+H = "/home/dev"; OUT = "/scratch/asr/data"; os.makedirs(OUT, exist_ok=True)
+random.seed(11)
+LIC = {"waxal": "CC-BY-SA-4.0 (WAXAL Media Trust subsets)", "fleurs": "CC-BY-4.0 (Google FLEURS)",
+       "slr70": "CC-BY-SA-4.0 (OpenSLR 70, Nigerian English)", "slr86": "CC-BY-SA-4.0 (OpenSLR 86, Yoruba)",
+       "bible_yor": "CC-BY-SA-4.0 (Open.Bible / Biblica)", "bible_hau": "CC-BY-SA-4.0 (Open.Bible / Biblica)",
+       "open_bible_igbo": "CC-BY-SA-4.0 (Open.Bible / Biblica)", "igboapi": "CC-BY-4.0 (IgboAPI)",
+       "naijas2st_igbo": "CC-BY-4.0 (NaijaS2ST)", "igbosyncorp": "CC0-1.0 (IgboSynCorp)"}
+# per-source sampling multiplier per epoch (Bible + dictionary reads down-weighted, Pidgin up-weighted)
+W = {"waxal:ibo": 1.0, "waxal:yor": 1.0, "waxal:hau": 1.0, "waxal:pcm": 3.0, "fleurs:eng": 1.0, "fleurs:hau": 1.0,
+     "fleurs:ibo": 1.0, "fleurs:yor": 1.0, "slr70:eng": 0.6, "slr86:yor": 1.0, "bible_yor:yor": 0.35, "bible_hau:hau": 0.35,
+     "open_bible_igbo:ibo": 0.15, "igboapi:ibo": 0.25, "naijas2st_igbo:ibo": 0.5, "igbosyncorp:ibo": 1.5}
+TONE = {"\u0300", "\u0301", "\u0304", "\u0302", "\u030c"}
+
+def norm(s, lang):
+    s = re.sub(r"<[^>]+>", " ", s); s = re.sub(r"\[\d+\]", " ", s)
+    s = unicodedata.normalize("NFC", s.lower())
+    if lang == "ibo":
+        s = unicodedata.normalize("NFC", "".join(c for c in unicodedata.normalize("NFD", s) if c not in TONE))
+    s = s.replace("\u2019", "'").replace("`", "'")
+    s = re.sub(r"[^\w\s']", " ", s); s = re.sub(r"(^|\s)'+|'+(\s|$)", " ", s)
+    return re.sub(r"\s+", " ", s).strip()
+
+def load(d):
+    return [json.loads(l) for l in open(f"{d}/meta.jsonl")]
+
+rows = []
+for ds in ["waxal", "fleurs", "slr70", "slr86", "bible_yor", "bible_hau"]:
+    d = f"{H}/data3/{ds}"
+    for r in load(d):
+        rows.append(dict(id=r["id"], src=ds, lang=r["lang"], spk=r.get("spk"), split=r.get("split"), raw=r["text"], dur=r["dur"],
+                         pcm=f"{d}/pcm24k.i16", off24=r["off24"], n24=r["n24"]))
+for ds in ["open_bible_igbo", "igboapi", "naijas2st_igbo", "igbosyncorp"]:
+    d = f"{H}/data_p3/p3set/{ds}"
+    for r in load(d):
+        rows.append(dict(id=r["id"], src=ds, lang=r["lang"], spk=r.get("speaker"), split=r.get("split"), raw=r.get("text_plain") or r.get("text_human"),
+                         dur=r.get("dur24") or r["dur"], pcm=f"{d}/pcm24k.i16", off24=r["off24"], n24=r["n24"]))
+for r in rows:
+    r["text"] = norm(r["raw"] or "", r["lang"]); r["licence"] = LIC[r["src"]]
+print("all rows", len(rows), collections.Counter((r["src"], r["split"]) for r in rows))
+
+# English lexicon (for detecting real English insertions in held-out Igbo/Yoruba/Hausa speech)
+ec, nc = collections.Counter(), collections.Counter()
+for r in rows:
+    if r["split"] != "train": continue
+    (ec if r["lang"] == "eng" else nc if r["lang"] in ("ibo", "yor", "hau") else collections.Counter()).update(r["text"].split())
+ENG = {w for w, c in ec.items() if c >= 2 and len(w) >= 3 and w.isascii() and w.isalpha() and nc[w] <= max(1, c // 50)}
+def eng_words(t):
+    return [w for w in t.split() if w in ENG]
+
+# real code-switch test: held-out ibo/yor/hau clips with >=1 English word (insertional CS, real speech) + IgboSynCorp natural CS
+realcs, hold_spk = [], set()
+for r in rows:
+    if r["lang"] not in ("ibo", "yor", "hau") or not r["text"]: continue
+    ew = eng_words(r["text"]); nw = len(r["text"].split())
+    if r["src"] == "igbosyncorp" and len(ew) >= 2 and nw >= 3:
+        hold_spk.add(r["spk"])
+    if r["split"] in ("test", "dev") and r["src"] in ("waxal", "naijas2st_igbo") and ew and nw >= 4 and len(ew) < nw:
+        realcs.append(dict(r, cs_eng=ew, cs_kind="insertional"))
+for r in rows:
+    if r["src"] == "igbosyncorp" and r["spk"] in hold_spk and eng_words(r["text"]) and len(r["text"].split()) >= 3:
+        realcs.append(dict(r, cs_eng=eng_words(r["text"]), cs_kind="natural_conversational"))
+print("realcs", len(realcs), collections.Counter((r["src"], r["lang"]) for r in realcs), "held-out igbosyncorp speakers", len(hold_spk))
+
+# dev: WAXAL validation 40/lang + FLEURS eng dev/test 40 (FLEURS val = the duplex eval English set, untouched here)
+dev = []
+for L in ["ibo", "yor", "hau", "pcm"]:
+    c = [r for r in rows if r["src"] == "waxal" and r["lang"] == L and r["split"] in ("validation", "val") and 1.5 < r["dur"] < 15]
+    random.shuffle(c); dev += c[:40]
+c = [r for r in rows if r["src"] == "slr70" and r["split"] != "train" and 1.5 < r["dur"] < 15]
+if not c:
+    sp = sorted({r["spk"] for r in rows if r["src"] == "slr70"}); hold70 = set(sp[:4])
+    c = [r for r in rows if r["src"] == "slr70" and r["spk"] in hold70]
+else:
+    hold70 = set()
+random.shuffle(c); dev += c[:40]
+
+train, drop = [], collections.Counter()
+for r in rows:
+    k = f"{r['src']}:{r['lang']}"
+    if r["split"] != "train": drop["not_train"] += 1; continue
+    if r["src"] == "slr70" and r["spk"] in hold70: drop["slr70_dev_spk"] += 1; continue
+    if r["src"] == "igbosyncorp" and r["spk"] in hold_spk: drop["realcs_spk"] += 1; continue
+    if k not in W: drop["no_weight:" + k] += 1; continue
+    if not (0.8 <= r["dur"] <= 20) or len(r["text"]) < 2: drop["dur_or_empty"] += 1; continue
+    if len(r["text"]) > r["dur"] * 25: drop["too_many_chars"] += 1; continue
+    r["w"] = W[k]; train.append(r)
+print("dropped", drop)
+def dump(name, R):
+    with open(f"{OUT}/{name}.jsonl", "w") as f:
+        for r in R:
+            f.write(json.dumps({k: v for k, v in r.items() if k != "raw"} | {"ref_raw": r["raw"]}, ensure_ascii=False) + "\n")
+dump("train", train); dump("dev", dev); dump("realcs_test", realcs)
+st = collections.defaultdict(lambda: [0, 0.0, 0.0])
+for r in train:
+    s = st[f"{r['src']}:{r['lang']}"]; s[0] += 1; s[1] += r["dur"] / 3600; s[2] += r["dur"] * r["w"] / 3600
+lang_h = collections.Counter()
+for k, v in st.items(): lang_h[k.split(":")[1]] += v[2]
+out = dict(sources={k: dict(n=v[0], hours=round(v[1], 2), hours_per_epoch=round(v[2], 2), licence=LIC[k.split(':')[0]]) for k, v in sorted(st.items())},
+           lang_hours_per_epoch={k: round(v, 1) for k, v in lang_h.items()}, total_hours=round(sum(v[1] for v in st.values()), 1),
+           epoch_hours=round(sum(v[2] for v in st.values()), 1), dev=len(dev), realcs=len(realcs), eng_lexicon=len(ENG))
+json.dump(out, open(f"{OUT}/stats.json", "w"), indent=1); print(json.dumps(out, indent=1))
+json.dump(sorted(ENG), open(f"{OUT}/eng_lexicon.json", "w"))
+print("PREP_DONE")

code/training/synth_cs.py

@@ -0,0 +1,62 @@
+"""Synthesize code-switched ASR training audio with Dynamic Accent Flash-DPO (public-safe, licensed-data TTS).
+~/venv/bin/python synth_cs.py <worker> <n_workers>  -> /scratch/asr/synth/wav/*.wav + parts/part_<k>.jsonl
+Voices: 1 train clip per WAXAL speaker (ibo/yor/hau 1-8, pcm F1-4/M1-4; CC-BY-SA-4.0 Media Trust) as zero-shot prompt + LM speaker id.
+Pace: random LM rate token (slow/normal/fast) + random speed perturbation 0.9-1.1. Channel/noise/reverb augmentation is applied on the fly in training."""
+import os, sys, json, random, re, unicodedata, time
+os.environ.setdefault("CV3_DIR", "/home/dev/models/Fun-CosyVoice3-0.5B-2512")
+REL = "/home/dev/rel_flash"; sys.path[:0] = [REL]
+import numpy as np, soundfile as sf, librosa, torch
+K, NW = int(sys.argv[1]), int(sys.argv[2])
+S_DIR = "/scratch/asr/synth"; os.makedirs(f"{S_DIR}/wav", exist_ok=True); os.makedirs(f"{S_DIR}/voices", exist_ok=True); os.makedirs(f"{S_DIR}/parts", exist_ok=True)
+import flash_lm
+TONE = {"\u0300", "\u0301", "\u0304", "\u0302", "\u030c"}
+def norm(s, lang):
+    s = unicodedata.normalize("NFC", s.lower())
+    if lang == "ibo": s = unicodedata.normalize("NFC", "".join(c for c in unicodedata.normalize("NFD", s) if c not in TONE))
+    s = s.replace("\u2019", "'"); s = re.sub(r"[^\w\s']", " ", s); s = re.sub(r"(^|\s)'+|'+(\s|$)", " ", s)
+    return re.sub(r"\s+", " ", s).strip()
+TAG = re.compile(r"<(eng|ibo|yor|hau|pcm)>(.*?)</\1>", re.S)
+
+# voice pool (deterministic, built by every worker identically; worker 0 writes the wavs)
+W = "/home/dev/data3/waxal"; meta = [json.loads(l) for l in open(f"{W}/meta.jsonl")]
+pcm = np.memmap(f"{W}/pcm24k.i16", dtype=np.int16, mode="r")
+rng0 = random.Random(3); by = {}
+for r in meta:
+    if r["split"] == "train" and 4.0 <= r["dur"] <= 9.0 and r.get("ascore", 0) > 0.4: by.setdefault(r["spk"], []).append(r)
+voices = {}
+for spk in sorted(by):
+    r = rng0.choice(by[spk]); p = f"{S_DIR}/voices/{spk}.wav"
+    if not os.path.exists(p):
+        tmp = p + f".{K}.tmp.wav"; sf.write(tmp, pcm[r["off24"]: r["off24"] + r["n24"]].astype(np.float32) / 32768, 24000); os.replace(tmp, p)
+    voices[spk] = dict(wav=p, spk=spk, lang=r["lang"], text=f"<{r['lang']}>{r['text']}</{r['lang']}>")
+for v, d in voices.items(): flash_lm.VOICES[v] = dict(wav=d["wav"], spk=d["spk"])
+S = flash_lm.Synth(f"{REL}/checkpoints/flash/flash_dpo/best.pt")
+for v, d in voices.items(): S.vtext[v] = d["text"]
+VL = sorted(voices); print("voices", len(VL), VL, flush=True)
+lines = [json.loads(l) for l in open(f"{S_DIR}/cs_text.jsonl")]
+mine = [l for i, l in enumerate(lines) if i % NW == K]
+out = open(f"{S_DIR}/parts/part_{K}.jsonl", "a"); done = {json.loads(l)["id"] for l in open(f"{S_DIR}/parts/part_{K}.jsonl")}
+rng = random.Random(100 + K); t0 = time.time(); n = 0
+for L in mine:
+    if L["id"] in done: continue
+    langs = L["pair"].split("+")
+    # prefer a voice whose native language is in the pair (natural accent), sometimes any voice (accent variety)
+    pool = [v for v in VL if voices[v]["lang"] in langs] if rng.random() < 0.75 else VL
+    v = rng.choice(pool or VL); rate = rng.choice(["slow", "normal", "normal", "normal", "fast"]); seed = rng.randrange(10 ** 6)
+    try:
+        toks = S.tokens(L["tagged"], v, rate, seed)[0]
+        if len(toks) < 10 or len(toks) > 25 * 30: continue
+        w = S.wav(toks, v); w = w.squeeze().detach().cpu().numpy() if torch.is_tensor(w) else np.asarray(w).squeeze()
+    except Exception as e:
+        print("ERR", L["id"], repr(e)[:200], flush=True); continue
+    y = librosa.resample(w.astype(np.float32), orig_sr=S.cosy.sample_rate, target_sr=16000)
+    sp = rng.uniform(0.9, 1.1)
+    if abs(sp - 1) > 0.02: y = librosa.resample(y, orig_sr=16000, target_sr=int(16000 / sp))   # speed perturb (played at 16 kHz)
+    y = y / (np.abs(y).max() + 1e-6) * rng.uniform(0.5, 0.95)
+    p = f"{S_DIR}/wav/{L['id']}.wav"; sf.write(p, y, 16000, subtype="PCM_16")
+    text = " ".join(norm(t, l) for l, t in TAG.findall(L["tagged"]))
+    out.write(json.dumps(dict(id=L["id"], wav=p, dur=round(len(y) / 16000, 2), lang="cs", src="synth_flash_dpo", text=text, tagged=L["tagged"],
+                              pair=L["pair"], voice=v, rate=rate, speed=round(sp, 3), licence="synthetic (Flash-DPO, licensed-data TTS)"), ensure_ascii=False) + "\n")
+    out.flush(); n += 1
+    if n % 20 == 0: print(f"w{K} {n}/{len(mine)} {(time.time() - t0) / n:.2f}s/clip", flush=True)
+print("SYNTH_DONE", K, n, flush=True)

code/training/train_asr.py

@@ -0,0 +1,185 @@
+"""Fine-tune Meta omniASR CTC (default omniASR_CTC_1B_v2, Apache-2.0) on Nigerian speech (+ optional synthetic code-switch).
+Full fine-tune (CNN feature extractor frozen), bf16 autocast, AdamW, warmup + linear decay, duration-bucketed batches.
+The model stays a plain CTC wav2vec2 (same architecture/tokenizer), so the duplex asr_server streams it unchanged.
+  ~/venv_omni/bin/python train_asr.py --run /scratch/asr/runs/A --steps 4000 [--synth /scratch/asr/synth/manifest.jsonl --synth_frac 0.2]
+"""
+import argparse, json, math, os, random, time, re, unicodedata
+import numpy as np, torch, torch.nn.functional as F
+import soundfile  # noqa: F401
+from scipy.signal import resample_poly, fftconvolve
+from fairseq2.models.hub import load_model
+from fairseq2.data.tokenizers.hub import load_tokenizer
+from fairseq2.nn.batch_layout import BatchLayout
+
+ap = argparse.ArgumentParser()
+ap.add_argument("--card", default="omniASR_CTC_1B_v2"); ap.add_argument("--run", required=True)
+ap.add_argument("--data", default="/scratch/asr/data"); ap.add_argument("--synth", default="")
+ap.add_argument("--synth_frac", type=float, default=0.2); ap.add_argument("--steps", type=int, default=4000)
+ap.add_argument("--lr", type=float, default=2e-5); ap.add_argument("--warmup", type=int, default=300)
+ap.add_argument("--batch_sec", type=float, default=280.0); ap.add_argument("--max_sec", type=float, default=20.0)
+ap.add_argument("--real_aug_p", type=float, default=0.15); ap.add_argument("--synth_aug_p", type=float, default=0.8)
+ap.add_argument("--val_every", type=int, default=500); ap.add_argument("--seed", type=int, default=1)
+ap.add_argument("--workers", type=int, default=10); ap.add_argument("--smoke", action="store_true")
+a = ap.parse_args()
+os.makedirs(a.run, exist_ok=True); json.dump(vars(a), open(f"{a.run}/args.json", "w"), indent=1)
+random.seed(a.seed); np.random.seed(a.seed); torch.manual_seed(a.seed)
+dev = torch.device("cuda")
+if os.environ.get("MEM_FRAC"): torch.cuda.set_per_process_memory_fraction(float(os.environ["MEM_FRAC"]))
+model = load_model(a.card, device=dev, dtype=torch.float32)
+tok = load_tokenizer(a.card); enc = tok.create_encoder(); dec = tok.create_decoder(skip_special_tokens=True)
+VI = tok.vocab_info; BLANK = 0; PAD = VI.pad_idx if VI.pad_idx is not None else 1
+print("vocab", VI.size, "pad", VI.pad_idx, "unk", VI.unk_idx, "bos", VI.bos_idx, "eos", VI.eos_idx, flush=True)
+SPECIAL = {i for i in (VI.pad_idx, VI.bos_idx, VI.eos_idx) if i is not None}
+
+def encode(t):
+    ids = [int(x) for x in enc(t).tolist()]
+    return [i for i in ids if i not in SPECIAL]
+
+def rows(p):
+    return [json.loads(l) for l in open(p)]
+train = rows(f"{a.data}/train.jsonl"); devset = rows(f"{a.data}/dev.jsonl")
+synth = rows(a.synth) if a.synth else []
+for r in train: r["kind"] = "real"
+for r in synth: r["kind"] = "synth"; r.setdefault("w", 1.0)
+unk = VI.unk_idx
+def ok(r):
+    ids = encode(r["text"]); r["n_ids"] = len(ids)
+    return len(ids) > 0 and (unk is None or unk not in ids) and r["dur"] <= a.max_sec
+n0 = len(train); train = [r for r in train if ok(r)]; synth = [r for r in synth if ok(r)]
+print(f"train {len(train)}/{n0} after tokenizer filter, synth {len(synth)}", flush=True)
+
+_mm = {}
+def load_audio(r):
+    if r.get("wav"):
+        y, sr = soundfile.read(r["wav"], dtype="float32")
+        if y.ndim > 1: y = y.mean(1)
+        if sr != 16000: y = resample_poly(y, 16000, sr).astype(np.float32)
+        return y
+    if r["pcm"] not in _mm: _mm[r["pcm"]] = np.memmap(r["pcm"], dtype=np.int16, mode="r")
+    y = _mm[r["pcm"]][r["off24"]: r["off24"] + r["n24"]].astype(np.float32) / 32768
+    return resample_poly(y, 2, 3).astype(np.float32)
+
+def augment(y, rng):
+    """noise (white/pink/babble-ish), simple synthetic reverb, phone band-limit, gain; ~ mic/channel variety"""
+    if rng.random() < 0.5:   # reverb: exponentially decaying noise RIR, RT60 0.15-0.7 s
+        rt = rng.uniform(0.15, 0.7); n = int(16000 * rt); t = np.arange(n) / 16000
+        rir = rng.standard_normal(n) * np.exp(-6.9 * t / rt); rir[0] = 1.0; rir /= np.sqrt((rir ** 2).sum())
+        y = fftconvolve(y, rir)[: len(y)].astype(np.float32)
+    if rng.random() < 0.35:  # telephone / cheap mic channel
+        y = resample_poly(resample_poly(y, 1, 2), 2, 1)[: len(y)].astype(np.float32) if rng.random() < 0.5 else \
+            resample_poly(resample_poly(y, 8, 16), 16, 8)[: len(y)].astype(np.float32)
+        if rng.random() < 0.5: y = np.tanh(y * rng.uniform(1.5, 4.0)).astype(np.float32)
+    if rng.random() < 0.7:   # additive noise at SNR 5-30 dB
+        n = rng.standard_normal(len(y)).astype(np.float32)
+        if rng.random() < 0.5: n = np.cumsum(n); n -= np.convolve(n, np.ones(400) / 400, "same")   # brownish
+        snr = rng.uniform(5, 30); ps = (y ** 2).mean() + 1e-8; pn = (n ** 2).mean() + 1e-8
+        y = y + n * np.sqrt(ps / (pn * 10 ** (snr / 10)))
+    y = y * rng.uniform(0.3, 1.5)
+    return np.clip(y, -1, 1).astype(np.float32)
+
+class DS(torch.utils.data.Dataset):
+    def __init__(self, items, train_mode): self.items, self.tm = items, train_mode
+    def __len__(self): return len(self.items)
+    def __getitem__(self, i):
+        r = self.items[i]; y = load_audio(r)
+        if self.tm:
+            rng = np.random.default_rng((i * 7919 + int(time.time() * 1000)) % 2 ** 32)
+            p = a.synth_aug_p if r["kind"] == "synth" else a.real_aug_p
+            if rng.random() < p: y = augment(y, rng)
+        return torch.from_numpy(y), torch.tensor(encode(r["text"])), i
+
+def epoch_order(rng):
+    """weighted sampling (source weights) + synth mixed at synth_frac of audio hours; bucketed by duration"""
+    pick = []
+    for r in train:
+        k = r["w"]; n = int(k) + (1 if rng.random() < k - int(k) else 0); pick += [r] * n
+    if synth:
+        real_h = sum(r["dur"] for r in pick); want = real_h * a.synth_frac / (1 - a.synth_frac); have = sum(r["dur"] for r in synth)
+        rep = want / have; s = []
+        for r in synth:
+            n = int(rep) + (1 if rng.random() < rep - int(rep) else 0); s += [r] * n
+        pick += s
+    rng.shuffle(pick)
+    batches = []
+    for c in range(0, len(pick), 4000):   # sort within chunks of 4000 to bucket by duration
+        ch = sorted(pick[c: c + 4000], key=lambda r: r["dur"]); cur, cs = [], 0.0
+        for r in ch:
+            if cs + r["dur"] > a.batch_sec and cur: batches.append(cur); cur, cs = [], 0.0
+            cur.append(r); cs += r["dur"]
+        if cur: batches.append(cur)
+    rng.shuffle(batches); return batches
+
+def collate(b):
+    ys = [F.layer_norm(x, x.shape) for x, _, _ in b]; L = torch.tensor([len(x) for x in ys])
+    X = torch.zeros(len(ys), int(L.max())); [X[i, : len(x)].copy_(x) for i, x in enumerate(ys)]
+    T = [t for _, t, _ in b]; TL = torch.tensor([len(t) for t in T]); Y = torch.cat(T)
+    return X, L, Y, TL
+
+def forward(X, L):
+    bl = BatchLayout(X.shape, seq_lens=L.tolist(), device=dev)
+    logits, blo = model(X, bl)
+    return logits, torch.tensor(blo.seq_lens, device=dev)
+
+for p in model.encoder_frontend.feature_extractor.parameters(): p.requires_grad = False
+params = [p for p in model.parameters() if p.requires_grad]
+print("trainable params %.1fM" % (sum(p.numel() for p in params) / 1e6), flush=True)
+opt = torch.optim.AdamW(params, lr=a.lr, betas=(0.9, 0.98), eps=1e-8, weight_decay=0.01)
+sched = torch.optim.lr_scheduler.LambdaLR(opt, lambda s: min(1.0, (s + 1) / a.warmup) * max(0.02, 1 - max(0, s - a.warmup) / max(1, a.steps - a.warmup)))
+
+def cer_eval(items, bs=16):
+    model.eval(); hyps = []
+    with torch.no_grad(), torch.autocast("cuda", dtype=torch.bfloat16):
+        for i in range(0, len(items), bs):
+            b = [(torch.from_numpy(load_audio(r)), torch.tensor([1]), 0) for r in items[i: i + bs]]
+            X, L, _, _ = collate(b); lg, ol = forward(X.to(dev), L)
+            pr = lg.argmax(-1)
+            for j in range(pr.shape[0]):
+                s = pr[j, : ol[j]]; m = torch.ones_like(s, dtype=torch.bool); m[1:] = s[1:] != s[:-1]; s = s[m]; s = s[s != BLANK]
+                hyps.append(str(dec(s)))
+    model.train()
+    import jiwer
+    def n(x): x = unicodedata.normalize("NFC", x.lower()); x = re.sub(r"[^\w\s']", " ", x); return re.sub(r"\s+", " ", x).strip() or "x"
+    out = {}
+    for lg in sorted({r["lang"] for r in items}):
+        idx = [k for k, r in enumerate(items) if r["lang"] == lg]
+        out[lg] = round(jiwer.cer([n(items[k]["text"]) for k in idx], [n(hyps[k]) for k in idx]), 4)
+    out["mean"] = round(float(np.mean([v for k, v in out.items()])), 4)
+    return out, hyps
+
+log = open(f"{a.run}/log.jsonl", "a")
+v0, _ = cer_eval(devset); print("dev step0", v0, flush=True); log.write(json.dumps(dict(step=0, dev=v0)) + "\n"); log.flush()
+best = v0["mean"]; step = 0; rng = random.Random(a.seed); model.train(); t0 = time.time(); ema = None
+steps_total = 60 if a.smoke else a.steps
+while step < steps_total:
+    batches = epoch_order(rng); print(f"epoch with {len(batches)} batches", flush=True)
+    lut = {}
+    flat = []
+    for b in batches:
+        flat.append([len(lut) + k for k in range(len(b))]); [lut.setdefault(len(lut), r) for r in b]
+    items = [lut[k] for k in range(len(lut))]
+    dl = torch.utils.data.DataLoader(DS(items, True), batch_sampler=flat, num_workers=a.workers, collate_fn=collate,
+                                     prefetch_factor=4, persistent_workers=False)
+    for X, L, Y, TL in dl:
+        X, Y = X.to(dev, non_blocking=True), Y.to(dev)
+        with torch.autocast("cuda", dtype=torch.bfloat16):
+            lg, ol = forward(X, L)
+        lp = F.log_softmax(lg.float(), -1).transpose(0, 1)
+        loss = F.ctc_loss(lp, Y, ol, TL.to(dev), blank=BLANK, reduction="sum", zero_infinity=True) / TL.sum()
+        opt.zero_grad(set_to_none=True); loss.backward()
+        gn = torch.nn.utils.clip_grad_norm_(params, 5.0)
+        if torch.isfinite(gn): opt.step()
+        sched.step(); step += 1
+        ema = loss.item() if ema is None else 0.98 * ema + 0.02 * loss.item()
+        if step % 25 == 0:
+            el = time.time() - t0
+            print(f"step {step} loss {ema:.4f} gn {gn:.2f} lr {sched.get_last_lr()[0]:.2e} {el / step:.2f}s/step mem {torch.cuda.max_memory_allocated() / 2**30:.1f}G", flush=True)
+            log.write(json.dumps(dict(step=step, loss=round(ema, 4), sps=round(el / step, 3))) + "\n"); log.flush()
+        if step % a.val_every == 0 or step == steps_total:
+            v, _ = cer_eval(devset); print("dev", step, v, flush=True); log.write(json.dumps(dict(step=step, dev=v)) + "\n"); log.flush()
+            sd = {k: t.to(torch.bfloat16) for k, t in model.state_dict().items()}
+            if v["mean"] < best and not a.smoke:
+                best = v["mean"]; torch.save(dict(model=sd, step=step, dev=v, card=a.card), f"{a.run}/best.pt"); print("saved best", flush=True)
+        if step >= steps_total: break
+if not a.smoke:
+    torch.save(dict(model={k: t.to(torch.bfloat16) for k, t in model.state_dict().items()}, step=step, card=a.card), f"{a.run}/last.pt")
+print("TRAIN_DONE best", best, flush=True)

codeswitch-experimental/config.json

@@ -0,0 +1,100 @@
+{
+ "base_model": "omniASR_CTC_1B_v2 (facebook omnilingual-asr, Apache-2.0)",
+ "model_card": "omniASR_CTC_1B_v2",
+ "architecture": "wav2vec2 CTC 1B (unchanged)",
+ "vocab_size": 10288,
+ "tokenizer": "omniASR_tokenizer_written_v2",
+ "checkpoint_format": "torch.save dict {model: state_dict (bf16), step, dev, card}",
+ "sample_rate": 16000,
+ "decoding": "greedy CTC, Latin-script logit mask, no language hint",
+ "languages": [
+  "en",
+  "pcm",
+  "ig",
+  "yo",
+  "ha"
+ ],
+ "license": "cc-by-sa-4.0",
+ "training": {
+  "steps": 4000,
+  "batch_audio_sec": 160,
+  "max_clip_sec": 18,
+  "lr": 2e-05,
+  "warmup": 300,
+  "schedule": "linear decay",
+  "optimizer": "AdamW b(0.9,0.98) wd 0.01",
+  "precision": "bf16 autocast",
+  "frozen": "CNN feature encoder",
+  "seed": 1,
+  "real_aug_p": 0.15,
+  "synthetic_codeswitch_fraction": 0.12,
+  "data": "same licensed real speech + synthetic code-switched audio (Dynamic Accent Flash-DPO TTS), 12% of sampled audio",
+  "synth_aug_p": 0.8
+ },
+ "name": "dynamic-asr/codeswitch-experimental",
+ "status": "experimental",
+ "file": "model.pt",
+ "scores": {
+  "duplex_eng": {
+   "n": 29,
+   "wer": 0.0922,
+   "cer": 0.0241,
+   "cer_nodiac": 0.0241
+  },
+  "duplex_pcm": {
+   "n": 40,
+   "wer": 0.1927,
+   "cer": 0.0761,
+   "cer_nodiac": 0.0759
+  },
+  "duplex_ibo": {
+   "n": 40,
+   "wer": 0.3207,
+   "cer": 0.095,
+   "cer_nodiac": 0.0754
+  },
+  "duplex_yor": {
+   "n": 40,
+   "wer": 0.3532,
+   "cer": 0.1182,
+   "cer_nodiac": 0.0782
+  },
+  "duplex_hau": {
+   "n": 40,
+   "wer": 0.252,
+   "cer": 0.0808,
+   "cer_nodiac": 0.0808
+  },
+  "duplex_cs": {
+   "n": 40,
+   "wer": 0.2157,
+   "cer": 0.0718,
+   "cer_nodiac": 0.0648
+  },
+  "realcs_all": {
+   "n": 94,
+   "wer": 0.438,
+   "cer": 0.1492,
+   "cer_nodiac": 0.1368
+  },
+  "realcs_natural_conversational": {
+   "n": 36,
+   "wer": 0.6103,
+   "cer": 0.2472,
+   "cer_nodiac": 0.2339
+  },
+  "realcs_insertional": {
+   "n": 58,
+   "wer": 0.3727,
+   "cer": 0.1186,
+   "cer_nodiac": 0.1065
+  },
+  "realcs_eng_word_recall": 0.3805,
+  "synthtest": {
+   "n": 109,
+   "wer": 0.1261,
+   "cer": 0.0385,
+   "cer_nodiac": 0.0349
+  }
+ }
+}

codeswitch-experimental/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7abf1f167e41f07af97c6ea261a218b7dbf931f7b18056167fa361fcb2cb05b8
+size 1951564394

config.json

@@ -0,0 +1,297 @@
+{
+ "name": "dynamic-asr",
+ "version": "0.1.0",
+ "pipeline_tag": "automatic-speech-recognition",
+ "library_name": "fairseq2 / omnilingual-asr",
+ "base_model": "facebook omniASR_CTC_1B_v2 (Apache-2.0)",
+ "architecture": "wav2vec2 CTC, 1B (unchanged; vocab 10288, tokenizer omniASR_tokenizer_written_v2)",
+ "languages": [
+  "en",
+  "pcm",
+  "ig",
+  "yo",
+  "ha"
+ ],
+ "default_model": "real",
+ "models": {
+  "real": {
+   "path": "real/model.pt",
+   "trained_on": "licensed real speech only",
+   "status": "default"
+  },
+  "codeswitch-experimental": {
+   "path": "codeswitch-experimental/model.pt",
+   "trained_on": "same real speech + synthetic code-switched audio (Flash-DPO TTS)",
+   "status": "experimental"
+  }
+ },
+ "decoding": {
+  "type": "greedy CTC",
+  "latin_script_mask": true,
+  "language_hint": null,
+  "streaming": "re-decode the growing buffer (duplex asr_server)"
+ },
+ "output": "lowercase, no punctuation; Igbo without tone marks (dot letters kept); Yoruba with tone marks",
+ "weights_license": "cc-by-sa-4.0",
+ "code_license": "apache-2.0",
+ "scores": {
+  "v1_stock": {
+   "duplex_eng": {
+    "n": 29,
+    "wer": 0.1437,
+    "cer": 0.0654,
+    "cer_nodiac": 0.0654
+   },
+   "duplex_pcm": {
+    "n": 40,
+    "wer": 0.3796,
+    "cer": 0.1693,
+    "cer_nodiac": 0.169
+   },
+   "duplex_ibo": {
+    "n": 40,
+    "wer": 0.636,
+    "cer": 0.2057,
+    "cer_nodiac": 0.1614
+   },
+   "duplex_yor": {
+    "n": 40,
+    "wer": 0.4978,
+    "cer": 0.1905,
+    "cer_nodiac": 0.1227
+   },
+   "duplex_hau": {
+    "n": 40,
+    "wer": 0.299,
+    "cer": 0.0897,
+    "cer_nodiac": 0.0885
+   },
+   "duplex_cs": {
+    "n": 40,
+    "wer": 0.3295,
+    "cer": 0.1282,
+    "cer_nodiac": 0.1053
+   },
+   "realcs_all": {
+    "n": 94,
+    "wer": 0.6062,
+    "cer": 0.2243,
+    "cer_nodiac": 0.1951
+   },
+   "realcs_natural_conversational": {
+    "n": 36,
+    "wer": 0.8351,
+    "cer": 0.3719,
+    "cer_nodiac": 0.3256
+   },
+   "realcs_insertional": {
+    "n": 58,
+    "wer": 0.5195,
+    "cer": 0.1783,
+    "cer_nodiac": 0.1544
+   },
+   "realcs_eng_word_recall": 0.3009,
+   "synthtest": {
+    "n": 109,
+    "wer": 0.4141,
+    "cer": 0.1586,
+    "cer_nodiac": 0.1293
+   }
+  },
+  "v2_stock": {
+   "duplex_eng": {
+    "n": 29,
+    "wer": 0.0969,
+    "cer": 0.0244,
+    "cer_nodiac": 0.0244
+   },
+   "duplex_pcm": {
+    "n": 40,
+    "wer": 0.3358,
+    "cer": 0.1232,
+    "cer_nodiac": 0.1223
+   },
+   "duplex_ibo": {
+    "n": 40,
+    "wer": 0.4559,
+    "cer": 0.1329,
+    "cer_nodiac": 0.1089
+   },
+   "duplex_yor": {
+    "n": 40,
+    "wer": 0.4009,
+    "cer": 0.1346,
+    "cer_nodiac": 0.0977
+   },
+   "duplex_hau": {
+    "n": 40,
+    "wer": 0.3172,
+    "cer": 0.1061,
+    "cer_nodiac": 0.1059
+   },
+   "duplex_cs": {
+    "n": 40,
+    "wer": 0.2922,
+    "cer": 0.1042,
+    "cer_nodiac": 0.0898
+   },
+   "realcs_all": {
+    "n": 94,
+    "wer": 0.5139,
+    "cer": 0.1931,
+    "cer_nodiac": 0.1725
+   },
+   "realcs_natural_conversational": {
+    "n": 36,
+    "wer": 0.7773,
+    "cer": 0.3535,
+    "cer_nodiac": 0.3152
+   },
+   "realcs_insertional": {
+    "n": 58,
+    "wer": 0.4141,
+    "cer": 0.1431,
+    "cer_nodiac": 0.1281
+   },
+   "realcs_eng_word_recall": 0.2832,
+   "synthtest": {
+    "n": 109,
+    "wer": 0.3541,
+    "cer": 0.1324,
+    "cer_nodiac": 0.1179
+   }
+  },
+  "A_real": {
+   "duplex_eng": {
+    "n": 29,
+    "wer": 0.0953,
+    "cer": 0.0252,
+    "cer_nodiac": 0.0252
+   },
+   "duplex_pcm": {
+    "n": 40,
+    "wer": 0.1927,
+    "cer": 0.0738,
+    "cer_nodiac": 0.0736
+   },
+   "duplex_ibo": {
+    "n": 40,
+    "wer": 0.3153,
+    "cer": 0.0929,
+    "cer_nodiac": 0.0761
+   },
+   "duplex_yor": {
+    "n": 40,
+    "wer": 0.3547,
+    "cer": 0.1213,
+    "cer_nodiac": 0.0841
+   },
+   "duplex_hau": {
+    "n": 40,
+    "wer": 0.2533,
+    "cer": 0.0823,
+    "cer_nodiac": 0.0823
+   },
+   "duplex_cs": {
+    "n": 40,
+    "wer": 0.2385,
+    "cer": 0.0807,
+    "cer_nodiac": 0.0646
+   },
+   "realcs_all": {
+    "n": 94,
+    "wer": 0.4357,
+    "cer": 0.1482,
+    "cer_nodiac": 0.137
+   },
+   "realcs_natural_conversational": {
+    "n": 36,
+    "wer": 0.6227,
+    "cer": 0.2481,
+    "cer_nodiac": 0.2335
+   },
+   "realcs_insertional": {
+    "n": 58,
+    "wer": 0.3648,
+    "cer": 0.117,
+    "cer_nodiac": 0.107
+   },
+   "realcs_eng_word_recall": 0.3186,
+   "synthtest": {
+    "n": 109,
+    "wer": 0.2932,
+    "cer": 0.1052,
+    "cer_nodiac": 0.0914
+   }
+  },
+  "B_synth": {
+   "duplex_eng": {
+    "n": 29,
+    "wer": 0.0922,
+    "cer": 0.0241,
+    "cer_nodiac": 0.0241
+   },
+   "duplex_pcm": {
+    "n": 40,
+    "wer": 0.1927,
+    "cer": 0.0761,
+    "cer_nodiac": 0.0759
+   },
+   "duplex_ibo": {
+    "n": 40,
+    "wer": 0.3207,
+    "cer": 0.095,
+    "cer_nodiac": 0.0754
+   },
+   "duplex_yor": {
+    "n": 40,
+    "wer": 0.3532,
+    "cer": 0.1182,
+    "cer_nodiac": 0.0782
+   },
+   "duplex_hau": {
+    "n": 40,
+    "wer": 0.252,
+    "cer": 0.0808,
+    "cer_nodiac": 0.0808
+   },
+   "duplex_cs": {
+    "n": 40,
+    "wer": 0.2157,
+    "cer": 0.0718,
+    "cer_nodiac": 0.0648
+   },
+   "realcs_all": {
+    "n": 94,
+    "wer": 0.438,
+    "cer": 0.1492,
+    "cer_nodiac": 0.1368
+   },
+   "realcs_natural_conversational": {
+    "n": 36,
+    "wer": 0.6103,
+    "cer": 0.2472,
+    "cer_nodiac": 0.2339
+   },
+   "realcs_insertional": {
+    "n": 58,
+    "wer": 0.3727,
+    "cer": 0.1186,
+    "cer_nodiac": 0.1065
+   },
+   "realcs_eng_word_recall": 0.3805,
+   "synthtest": {
+    "n": 109,
+    "wer": 0.1261,
+    "cer": 0.0385,
+    "cer_nodiac": 0.0349
+   }
+  }
+ },
+ "excluded_data": [
+  "African Voices / WaZoBia",
+  "NaijaVoices (non-commercial)",
+  "AfriSpeech-200 (non-commercial)",
+  "Common Voice (not fetched)"
+ ]
+}

real/config.json

@@ -0,0 +1,99 @@
+{
+ "base_model": "omniASR_CTC_1B_v2 (facebook omnilingual-asr, Apache-2.0)",
+ "model_card": "omniASR_CTC_1B_v2",
+ "architecture": "wav2vec2 CTC 1B (unchanged)",
+ "vocab_size": 10288,
+ "tokenizer": "omniASR_tokenizer_written_v2",
+ "checkpoint_format": "torch.save dict {model: state_dict (bf16), step, dev, card}",
+ "sample_rate": 16000,
+ "decoding": "greedy CTC, Latin-script logit mask, no language hint",
+ "languages": [
+  "en",
+  "pcm",
+  "ig",
+  "yo",
+  "ha"
+ ],
+ "license": "cc-by-sa-4.0",
+ "training": {
+  "steps": 4000,
+  "batch_audio_sec": 160,
+  "max_clip_sec": 18,
+  "lr": 2e-05,
+  "warmup": 300,
+  "schedule": "linear decay",
+  "optimizer": "AdamW b(0.9,0.98) wd 0.01",
+  "precision": "bf16 autocast",
+  "frozen": "CNN feature encoder",
+  "seed": 1,
+  "real_aug_p": 0.15,
+  "synthetic_codeswitch_fraction": 0.0,
+  "data": "licensed real Nigerian speech only"
+ },
+ "name": "dynamic-asr/real",
+ "status": "default",
+ "file": "model.pt",
+ "scores": {
+  "duplex_eng": {
+   "n": 29,
+   "wer": 0.0953,
+   "cer": 0.0252,
+   "cer_nodiac": 0.0252
+  },
+  "duplex_pcm": {
+   "n": 40,
+   "wer": 0.1927,
+   "cer": 0.0738,
+   "cer_nodiac": 0.0736
+  },
+  "duplex_ibo": {
+   "n": 40,
+   "wer": 0.3153,
+   "cer": 0.0929,
+   "cer_nodiac": 0.0761
+  },
+  "duplex_yor": {
+   "n": 40,
+   "wer": 0.3547,
+   "cer": 0.1213,
+   "cer_nodiac": 0.0841
+  },
+  "duplex_hau": {
+   "n": 40,
+   "wer": 0.2533,
+   "cer": 0.0823,
+   "cer_nodiac": 0.0823
+  },
+  "duplex_cs": {
+   "n": 40,
+   "wer": 0.2385,
+   "cer": 0.0807,
+   "cer_nodiac": 0.0646
+  },
+  "realcs_all": {
+   "n": 94,
+   "wer": 0.4357,
+   "cer": 0.1482,
+   "cer_nodiac": 0.137
+  },
+  "realcs_natural_conversational": {
+   "n": 36,
+   "wer": 0.6227,
+   "cer": 0.2481,
+   "cer_nodiac": 0.2335
+  },
+  "realcs_insertional": {
+   "n": 58,
+   "wer": 0.3648,
+   "cer": 0.117,
+   "cer_nodiac": 0.107
+  },
+  "realcs_eng_word_recall": 0.3186,
+  "synthtest": {
+   "n": 109,
+   "wer": 0.2932,
+   "cer": 0.1052,
+   "cer_nodiac": 0.0914
+  }
+ }
+}

real/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ab32b443451935ec12d1cf836b7433e2057cb0f9f95913e1bf9f1a64a628b5c
+size 1951564394

transcribe.py

@@ -0,0 +1,65 @@
+"""Dynamic ASR: transcribe Nigerian English, Pidgin, Igbo, Yoruba, Hausa and code-switched speech.
+
+  pip install "omnilingual-asr>=0.2.0" huggingface_hub soundfile scipy
+  python transcribe.py audio.wav [more.wav ...] [--model real|codeswitch] [--no-latin-mask]
+
+Model: Meta omniASR_CTC_1B_v2 (Apache-2.0) fine-tuned by Modular. Greedy CTC decoding, no language hint needed.
+Output: lowercase text without punctuation (Igbo without tone marks; Yoruba with tone marks).
+"""
+import argparse, unicodedata
+import numpy as np, torch, torch.nn.functional as F
+import soundfile as sf
+from scipy.signal import resample_poly
+from huggingface_hub import hf_hub_download
+from omnilingual_asr.models.inference.pipeline import ASRInferencePipeline  # noqa: F401 (registers omniASR cards)
+from fairseq2.models.hub import load_model
+from fairseq2.data.tokenizers.hub import load_tokenizer
+from fairseq2.nn.batch_layout import BatchLayout
+
+REPO = "Modularcomputing/dynamic-asr"
+FILES = {"real": "real/model.pt", "codeswitch": "codeswitch-experimental/model.pt"}
+BASE = "omniASR_CTC_1B_v2"
+
+
+def latin_ok(s):
+    for c in s:
+        cat = unicodedata.category(c)
+        if cat[0] == "M":
+            if not ("\u0300" <= c <= "\u036f"):
+                return False
+        elif c.isalpha() and "LATIN" not in unicodedata.name(c, ""):
+            return False
+    return True
+
+
+class DynamicASR:
+    def __init__(self, model="real", device="cuda" if torch.cuda.is_available() else "cpu", latin_mask=True, ckpt=None):
+        self.dev = torch.device(device)
+        dtype = torch.bfloat16 if self.dev.type == "cuda" else torch.float32
+        self.model = load_model(BASE, device=self.dev, dtype=dtype).eval()   # base weights, then overwritten
+        ck = torch.load(ckpt or hf_hub_download(REPO, FILES[model]), map_location="cpu", weights_only=False)
+        self.model.load_state_dict({k: v.to(dtype) for k, v in ck["model"].items()})
+        self.dtype = dtype
+        self.dec = load_tokenizer(BASE).create_decoder(skip_special_tokens=True)
+        V = self.model.final_proj.weight.shape[0]
+        self.bad = torch.tensor([i for i in range(V) if not latin_ok(str(self.dec(torch.tensor([i]))))], device=self.dev) if latin_mask else None
+
+    @torch.no_grad()
+    def __call__(self, wav, sr=16000):
+        y = np.asarray(wav, dtype=np.float32)
+        if y.ndim > 1: y = y.mean(1)
+        if sr != 16000: y = resample_poly(y, 16000, sr).astype(np.float32)
+        x = torch.from_numpy(y); x = F.layer_norm(x, x.shape)[None].to(self.dev, self.dtype)
+        logits, lay = self.model(x, BatchLayout(x.shape, seq_lens=[x.shape[1]], device=self.dev))
+        if self.bad is not None: logits[..., self.bad] = -float("inf")
+        s = logits[0, : lay.seq_lens[0]].argmax(-1); m = torch.ones_like(s, dtype=torch.bool); m[1:] = s[1:] != s[:-1]
+        return str(self.dec(s[m]))
+
+
+if __name__ == "__main__":
+    ap = argparse.ArgumentParser(); ap.add_argument("audio", nargs="+"); ap.add_argument("--model", default="real", choices=list(FILES))
+    ap.add_argument("--no-latin-mask", action="store_true"); ap.add_argument("--ckpt", default=None, help="local model.pt instead of the Hub")
+    a = ap.parse_args()
+    asr = DynamicASR(a.model, latin_mask=not a.no_latin_mask, ckpt=a.ckpt)
+    for f in a.audio:
+        w, sr = sf.read(f, dtype="float32"); print(f"{f}\t{asr(w, sr)}")