Initial release

.gitattributes

@@ -33,3 +33,12 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+architecture.png filter=lfs diff=lfs merge=lfs -text
+examples/commonvoice_1.wav filter=lfs diff=lfs merge=lfs -text
+examples/commonvoice_2.wav filter=lfs diff=lfs merge=lfs -text
+examples/commonvoice_3.wav filter=lfs diff=lfs merge=lfs -text
+examples/commonvoice_4.wav filter=lfs diff=lfs merge=lfs -text
+examples/commonvoice_5.wav filter=lfs diff=lfs merge=lfs -text
+examples/commonvoice_6.wav filter=lfs diff=lfs merge=lfs -text
+examples/voxpopuli_1.wav filter=lfs diff=lfs merge=lfs -text
+examples/voxpopuli_2.wav filter=lfs diff=lfs merge=lfs -text

LICENSE

@@ -0,0 +1,408 @@
+Attribution-NonCommercial 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-NonCommercial 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-NonCommercial 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. 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.
+  d. 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.
+
+  e. 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.
+
+  f. Licensed Material means the artistic or literary work, database,
+     or other material to which the Licensor applied this Public
+     License.
+
+  g. 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.
+
+  h. Licensor means the individual(s) or entity(ies) granting rights
+     under this Public License.
+
+  i. NonCommercial means not primarily intended for or directed towards
+     commercial advantage or monetary compensation. For purposes of
+     this Public License, the exchange of the Licensed Material for
+     other material subject to Copyright and Similar Rights by digital
+     file-sharing or similar means is NonCommercial provided there is
+     no payment of monetary compensation in connection with the
+     exchange.
+
+  j. 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.
+
+  k. 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.
+
+  l. 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, for NonCommercial purposes only; and
+
+            b. produce, reproduce, and Share Adapted Material for
+               NonCommercial purposes only.
+
+       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. 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, including when
+          the Licensed Material is used other than for NonCommercial
+          purposes.
+
+
+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.
+
+       4. If You Share Adapted Material You produce, the Adapter's
+          License You apply must not prevent recipients of the Adapted
+          Material from complying with this Public License.
+
+
+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 for NonCommercial purposes
+     only;
+
+  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; 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.
+

README.md

@@ -0,0 +1,184 @@
+---
+language:
+- fi
+license: cc-by-nc-4.0
+library_name: pytorch
+pipeline_tag: automatic-speech-recognition
+tags:
+- speech-recognition
+- streaming
+- transducer
+- rnnt
+- mingru
+- finnish
+model-index:
+- name: mateli
+  results:
+  - task:
+      type: automatic-speech-recognition
+    dataset:
+      name: Common Voice 17.0 (fi, test)
+      type: mozilla-foundation/common_voice_17_0
+      config: fi
+      split: test
+    metrics:
+    - type: wer
+      value: 15.03
+    - type: cer
+      value: 3.08
+  - task:
+      type: automatic-speech-recognition
+    dataset:
+      name: FLEURS (fi, SIB-Fleurs subset)
+      type: google/fleurs
+      config: fi_fi
+    metrics:
+    - type: wer
+      value: 18.47
+    - type: cer
+      value: 7.49
+  - task:
+      type: automatic-speech-recognition
+    dataset:
+      name: VoxPopuli (fi, test)
+      type: facebook/voxpopuli
+      config: fi
+      split: test
+    metrics:
+    - type: wer
+      value: 17.44
+    - type: cer
+      value: 7.62
+---
+
+# Mateli: streaming Finnish speech recognition
+
+Mateli is a small streaming speech recognition model for Finnish (59.6 M parameters). Audio is processed in 320 ms chunks, each chunk sees only itself and the past, and the state carried between chunks has a fixed size. That makes it cheap enough to run live on a phone. This repository has the PyTorch weights and a single-file reference implementation. For streaming on phones and CPUs, use the ONNX export, [samuli/mateli-onnx](https://huggingface.co/samuli/mateli-onnx).
+
+Mateli is not the most accurate Finnish recogniser. It makes mistakes on every kind of speech, and more of them on spontaneous conversation than on read or prepared speech. Its strengths are streaming output, low latency and a low cost to run.
+
+Its output is lowercase words with word times, without punctuation. [samuli/mateli-punct](https://huggingface.co/samuli/mateli-punct) adds punctuation and capitals.
+
+## Examples
+
+Eight clips from public test sets are in `examples/`, and `examples/references.tsv` lists their sources and reference transcripts. The outputs below are from `python mateli.py examples/<clip>.wav`.
+
+Sentences written in colloquial Finnish (Common Voice), with [whisper-large-v3](https://huggingface.co/openai/whisper-large-v3) for comparison:
+
+| clip | reference | whisper-large-v3 | Mateli |
+|---|---|---|---|
+| `commonvoice_4` | mä en oo ehtiny vielä katsoa. | Mä en ole ehtinyt vielä katsoa. | mä en oo ehtiny vielä katsoa |
+| `commonvoice_5` | miten me voidaa olla varmoja et sä puhut totta? | Miten me voidaan olla varmoja, että sä puhut totta? | miten me voidaan olla varmoja et sä puhu totta |
+| `commonvoice_6` | sano vaa aika ja paikka ni miä tuun mukkaa kuulin itseni sanovan. | Sano vaan aika ja paikka, niin minä tuun mukaan, kuulin itseni sanoa. | sano vaa aika ja paikka niin minä tuun mukkaa kuulin itseni sanovan |
+
+Mateli writes the words in the form they were spoken (*oo ehtiny*, *et*, *vaa*, *mukkaa*), while Whisper turns them into standard written Finnish (*ole ehtinyt*, *että*, *vaan*, *mukaan*). Mateli also makes mistakes here. It writes *voidaan* for *voidaa*, *puhu* for *puhut* and *niin minä* for *ni miä*. Whisper was run with faster-whisper (int8 on CPU, beam 5, language set to Finnish).
+
+Sentences in written Finnish (Common Voice and VoxPopuli):
+
+| clip | reference | output |
+|---|---|---|
+| `commonvoice_1` | äänestin päätöslauselmaa vastaan koska siinä ei käsitellä riittävästi näitä seikkoja. | äänestin päätöslauselmaa vastaan koska siinä ei käsitellä riittävästi näitä seikkoja |
+| `commonvoice_2` | ne koostuivat eritavoin asetelluista ympyröistä. | ne koostuivat eri tavoin asetelluissa ympyröistä |
+| `commonvoice_3` | kaikki mikä tekisi tästä tuskasta edes hitusen siedettävämpää oli kiellettyä. | kaikki mikä tekis tästä tuskasta edes hitusen siedettävämpää oli kiellettyä |
+| `voxpopuli_1` | arvoisa puhemies äänestin tämän mietinnön puolesta. vaikka komission tiedonannosta on kulunut jo yli kymmenen vuotta eivät kaikki matkustajat edelleenkään tiedä oikeuksistaan. | repo äänestin tämän mietinnön puolesta öö vaikka komission tiedonainnosta on kulunut jo yli kymmeneen vuotta eivät kaikki matkustajat edelleenkään tiedä oikeuksistaan |
+| `voxpopuli_2` | boikotoiminen vaaleissa ja mielenosoitus ovat eri asioita mutta erittäin huolestuttavia ovat tiedot siitä että useampi kuin joka kymmenes äänestyspaikka ei voinut avata oviaan opposition saarron takia. | boiko toiminen vaaleissa ja mielenosoitus ovat eri asioita mutta erittäin huolestuttavia ovat tiedot siitä että useampi kuin joka kymmenen säänestyspaikka ei voinut avata oviaan opposition saannon takia |
+
+Typical mistakes are wrong case endings (*asetelluissa*), split or misheard words (*boiko toiminen*, *säänestyspaikka*), an unclear start turned into a different word (*repo* for *arvoisa puhemies*), hesitation sounds written out (*öö*) and spoken forms kept as spoken (*tekis*).
+
+Word times with `--words` (start and end in seconds):
+
+```
+$ python mateli.py examples/commonvoice_2.wav --words
+   1.60    1.64  ne
+   1.76    2.40  koostuivat
+   2.60    2.72  eri
+   2.84    3.16  tavoin
+   3.28    4.04  asetelluissa
+   4.20    4.72  ympyröistä
+```
+
+## How to use
+
+```
+pip install torch safetensors soundfile scipy
+hf download samuli/mateli --local-dir mateli
+python mateli/mateli.py mateli/examples/commonvoice_2.wav
+```
+
+From Python (run next to the downloaded `mateli` directory):
+
+```python
+import sys
+sys.path.insert(0, "mateli")
+from mateli import Mateli, load_audio
+
+asr = Mateli.from_pretrained("mateli")          # GPU if available, else CPU
+result = asr.transcribe(load_audio("mateli/examples/commonvoice_3.wav"))
+print(result.text)
+for word in result.words[:3]:
+    print(word)
+```
+
+```
+kaikki mikä tekis tästä tuskasta edes hitusen siedettävämpää oli kiellettyä
+Word(text='kaikki', start=1.16, end=1.44)
+Word(text='mikä', start=1.72, end=1.88)
+Word(text='tekis', start=1.96, end=2.12)
+```
+
+`transcribe()` takes float32 mono 16 kHz samples and runs the same 320 ms chunk loop that live streaming would, so its output equals streaming recognition of the same audio, and memory use stays flat on long recordings. The output is the same as the ONNX export's fp32 encoder.
+
+### Files
+
+| file | what |
+|---|---|
+| `model.safetensors` | weights, fp32 (238 MB) |
+| `config.json` | model settings and vocabulary |
+| `mateli.py` | model definition, transcription, command line |
+
+## Model
+
+![Mateli architecture and streaming](architecture.png)
+
+| | |
+|---|---|
+| parameters | 59.6 M |
+| input | 16 kHz mono, 80-bin log-mel (25 ms window, 10 ms hop) |
+| encoder | two stride-2 Conv2d (40 ms frames), 16 blocks of feed-forward / MinGRU / causal depthwise convolution / feed-forward, windowed self-attention after blocks 4, 7, 10 and 13 |
+| attention window | the current 320 ms chunk + 64 earlier frames (2.56 s), no positional encoding |
+| decoder | RNN-T: one-layer MinGRU prediction network, additive joiner, greedy search |
+| output | 44 characters (blank, space, a-z, å, ä, ö, 0-9, `. ? !`) |
+| latency | 320 ms chunks |
+| compute | about 2 GMAC per second of audio |
+
+[MinGRU](https://arxiv.org/abs/2410.01201) is a minimal gated recurrent unit whose gates do not depend on the previous state, so a whole chunk is computed in parallel with a log-space scan while streaming still carries a single state vector per layer. The model was trained with the RNN-T loss and an auxiliary CTC loss (the CTC head is not included).
+
+## Training data
+
+About 2,600 hours of Finnish speech from various publicly available sources and different registers. Transcripts are either human produced or from larger Finnish speech recognition models. Most of the machine transcripts were produced with [GetmanY1/wav2vec2-xlarge-fi-150k-finetuned](https://huggingface.co/GetmanY1/wav2vec2-xlarge-fi-150k-finetuned).
+
+The transcripts of conversational speech are deliberately verbatim. They write down what was said, in the form it was said (*mä*, *tekis*, *mahollisimman*), including hesitations like *öö*, instead of turning it into written language. Spoken Finnish differs a lot from written Finnish, and a verbatim transcript keeps that information. It stays faithful to the speaker, every spoken word gets its own time, and written forms can still be produced afterwards. Words that a cleaned-up transcript dropped or rewrote cannot be recovered. Read and prepared speech was transcribed in standard written form. The model learned both styles, so its output follows the speech. The cost is a higher word error rate against written-language references (see [Evaluation](#evaluation)).
+
+## Evaluation
+
+Word and character error rates (%). References and outputs are lowercased, with punctuation removed and foreign diacritics removed (é becomes e).
+
+| test set | clips | hours | words | WER | CER |
+|---|---:|---:|---:|---:|---:|
+| Common Voice 17.0, fi test (read sentences) | 1,763 | 2.50 | 11,135 | 15.03 | 3.08 |
+| FLEURS fi, SIB-Fleurs subset (read sentences) | 175 | 0.64 | 2,766 | 18.47 | 7.49 |
+| VoxPopuli fi test (parliamentary speeches) | 452 | 1.27 | 7,338 | 17.44 | 7.62 |
+
+Part of every WER comes from spelling choices. Spoken forms (*mä*, *tekis*) and compound words written apart or together count as errors against a written-language reference.
+
+## Limitations
+
+- No punctuation or capitals, and numbers are written as words. Use [samuli/mateli-punct](https://huggingface.co/samuli/mateli-punct) for punctuation and casing.
+- Accuracy drops on spontaneous conversation, overlapping speakers, dialects, noise and rare names.
+- Greedy decoding only, no language model.
+- Finnish only.
+- Behaviour on music, other languages and long non-speech passages has not been evaluated.
+
+## License
+
+The model weights and the code in this repository are licensed under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/). The example clips in `examples/` are CC0 (Common Voice and VoxPopuli).

config.json

@@ -0,0 +1,20 @@
+{
+  "model_type": "mateli-rnnt",
+  "sample_rate": 16000,
+  "n_mels": 80,
+  "d_model": 384,
+  "n_blocks": 16,
+  "conv_kernel": 15,
+  "ffn_expansion": 4,
+  "attn_after_blocks": [4, 7, 10, 13],
+  "attn_heads": 6,
+  "attn_left_context": 64,
+  "chunk_mel_frames": 32,
+  "frame_seconds": 0.04,
+  "pred_hidden": 384,
+  "joint_dim": 384,
+  "vocab": ["<blank>", " ", "a", "b", "c", "d", "e", "f", "g", "h", "i", "j", "k", "l", "m", "n", "o", "p", "q", "r", "s", "t", "u", "v", "w", "x", "y", "z", "å", "ä", "ö", "0", "1", "2", "3", "4", "5", "6", "7", "8", "9", ".", "?", "!"],
+  "blank_id": 0,
+  "space_id": 1,
+  "max_symbols_per_frame": 8
+}

examples/commonvoice_1.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c6d67d2fe3d216f38da55485c6cb385aa91ca2f58070fefc0f1030d2c6cc159f
+size 202796

examples/commonvoice_2.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:33afa4226a5e70f322616e1fb8f91af43881ff5f9664c597c64999e0911dbb3b
+size 193580

examples/commonvoice_3.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:93496eaca60839a576bfd8ed4472f158b4980b9dbaab81d0a9b5c59bfdf1aa6a
+size 203948

examples/commonvoice_4.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:20345a545034417533fa5c15246f294865fb4afe3eb692ab74d512247070b315
+size 114476

examples/commonvoice_5.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:735719b29eee996c2453a59275a9bb3e152fb77588bc980f1c68d122db4b3371
+size 171692

examples/commonvoice_6.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:32adf4f3aef9cd037aef55a350c4fa3ec821e22c26db3d9a1679c0a2a701383c
+size 218924

examples/references.tsv

@@ -0,0 +1,9 @@
+file	source	clip	licence	reference
+commonvoice_1.wav	Common Voice 17.0, Finnish test split	common_voice_fi_32490577.mp3	CC0-1.0	äänestin päätöslauselmaa vastaan koska siinä ei käsitellä riittävästi näitä seikkoja.
+commonvoice_2.wav	Common Voice 17.0, Finnish test split	common_voice_fi_25005783.mp3	CC0-1.0	ne koostuivat eritavoin asetelluista ympyröistä.
+commonvoice_3.wav	Common Voice 17.0, Finnish test split	common_voice_fi_25260168.mp3	CC0-1.0	kaikki mikä tekisi tästä tuskasta edes hitusen siedettävämpää oli kiellettyä.
+commonvoice_4.wav	Common Voice 17.0, Finnish test split	common_voice_fi_23716223.mp3	CC0-1.0	mä en oo ehtiny vielä katsoa.
+commonvoice_5.wav	Common Voice 17.0, Finnish test split	common_voice_fi_25132818.mp3	CC0-1.0	miten me voidaa olla varmoja et sä puhut totta?
+commonvoice_6.wav	Common Voice 17.0, Finnish test split	common_voice_fi_25018016.mp3	CC0-1.0	sano vaa aika ja paikka ni miä tuun mukkaa kuulin itseni sanovan.
+voxpopuli_1.wav	VoxPopuli, Finnish transcribed test split	fi_20121023-19:15:14_0.opus	CC0-1.0	arvoisa puhemies äänestin tämän mietinnön puolesta. vaikka komission tiedonannosta on kulunut jo yli kymmenen vuotta eivät kaikki matkustajat edelleenkään tiedä oikeuksistaan.
+voxpopuli_2.wav	VoxPopuli, Finnish transcribed test split	fi_20140206-15:12:12_3.opus	CC0-1.0	boikotoiminen vaaleissa ja mielenosoitus ovat eri asioita mutta erittäin huolestuttavia ovat tiedot siitä että useampi kuin joka kymmenes äänestyspaikka ei voinut avata oviaan opposition saarron takia.

examples/voxpopuli_1.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:09205953b9cc84e7a76a42f69bfebd8b15a4b10a0aa6b8d4a0195750dff05492
+size 398764

examples/voxpopuli_2.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:44eca13281e0d8a3b1968f3db6aa0148fbcd91dcb6fe2fd00c8319b5f875c437
+size 479338

mateli.py

@@ -0,0 +1,352 @@
+"""Mateli: streaming Finnish speech recognition (MinGRU encoder + RNN-T), PyTorch reference.
+
+    python mateli.py speech.wav                  # text
+    python mateli.py speech.wav --words          # one word per line with start/end times
+
+    from mateli import Mateli, load_audio
+    asr = Mateli.from_pretrained(".")            # directory with config.json and model.safetensors
+    result = asr.transcribe(load_audio("speech.wav"))
+    print(result.text)
+
+The encoder is chunk-causal: audio is processed in 320 ms chunks (32 log-mel frames -> 8 encoder
+frames of 40 ms), each chunk sees only itself and the past, and the state carried between chunks
+has a fixed size. transcribe() runs the same chunk loop a live stream would, so its output equals
+streaming recognition of the same audio.
+
+Model: log-mel (80 bins, 25 ms window, 10 ms hop) -> two stride-2 Conv2d -> 16 macaron blocks
+(feed-forward / MinGRU / causal depthwise conv / feed-forward), with a windowed self-attention
+sublayer after blocks 4, 7, 10 and 13 (keys = the current chunk + 64 previous frames) -> additive
+joiner with a one-layer MinGRU prediction network over 44 characters. Greedy decoding.
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import math
+from dataclasses import dataclass
+from pathlib import Path
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+# --------------------------------------------------------------------------- MinGRU
+class MinGRU(nn.Module):
+    """minGRU (Feng et al. 2024, "Were RNNs All We Needed?") with a sigmoid highway.
+    h_t = (1 - z_t) * h_{t-1} + z_t * g(x_t), computed for a whole chunk with a log-space scan."""
+
+    def __init__(self, hidden: int):
+        super().__init__()
+        self.layers = nn.ModuleList([nn.Linear(hidden, 3 * hidden, bias=False)])
+
+    @staticmethod
+    def _log_g(x):
+        return torch.where(x >= 0, (F.relu(x) + 0.5).log(), -F.softplus(-x))
+
+    def forward(self, x, state):
+        """x [B, T, H], state [B, H] (>= 0) -> (output [B, T, H], new state [B, H])."""
+        hidden, gate, proj = self.layers[0](x).chunk(3, dim=-1)
+        log_coeffs = -F.softplus(gate)
+        log_values = -F.softplus(-gate) + self._log_g(hidden)
+        # the carried state enters the scan as an extra first step with coefficient 1
+        log_coeffs = torch.cat([torch.zeros_like(log_coeffs[:, :1]), log_coeffs], dim=1)
+        log_values = torch.cat([state.clamp_min(0).log().unsqueeze(1), log_values], dim=1)
+        a_star = log_coeffs.cumsum(dim=1)
+        h = (a_star + (log_values - a_star).logcumsumexp(dim=1)).exp()[:, 1:]
+        g = proj.sigmoid()
+        return g * h + (1.0 - g) * x, h[:, -1]
+
+    def step(self, x, state):
+        """One time step without the scan: x [B, H], state [B, H] -> (output, new state)."""
+        hidden, gate, proj = self.layers[0](x).chunk(3, dim=-1)
+        g_hidden = torch.where(hidden >= 0, hidden + 0.5, hidden.sigmoid())
+        h = torch.lerp(state, g_hidden, gate.sigmoid())
+        g = proj.sigmoid()
+        return g * h + (1.0 - g) * x, h
+
+
+# --------------------------------------------------------------------------- encoder
+class FeedForward(nn.Module):
+    def __init__(self, d: int, expansion: int):
+        super().__init__()
+        self.net = nn.Sequential(nn.LayerNorm(d), nn.Linear(d, expansion * d), nn.SiLU(),
+                                 nn.Linear(expansion * d, d))
+
+    def forward(self, x):
+        return self.net(x)
+
+
+class MinGRUMixer(nn.Module):
+    def __init__(self, d: int):
+        super().__init__()
+        self.norm = nn.LayerNorm(d)
+        self.gru = MinGRU(d)
+
+    def forward(self, x, state):
+        return self.gru(self.norm(x), state)
+
+
+class ConvModule(nn.Module):
+    """Conformer convolution module with a causal depthwise convolution."""
+
+    def __init__(self, d: int, kernel: int):
+        super().__init__()
+        self.norm = nn.LayerNorm(d)
+        self.pw1 = nn.Conv1d(d, 2 * d, 1)
+        self.dw = nn.Conv1d(d, d, kernel, groups=d)
+        self.act_norm = nn.LayerNorm(d)
+        self.pw2 = nn.Conv1d(d, d, 1)
+
+    def forward(self, x, cache):
+        """x [B, T, D], cache [B, D, kernel-1] (the previous GLU outputs)."""
+        y = F.glu(self.pw1(self.norm(x).transpose(1, 2)), dim=1)
+        y = torch.cat([cache, y], dim=2)
+        new_cache = y[:, :, -cache.shape[2]:]
+        y = F.silu(self.act_norm(self.dw(y).transpose(1, 2)))
+        return self.pw2(y.transpose(1, 2)).transpose(1, 2), new_cache
+
+
+class Block(nn.Module):
+    def __init__(self, d: int, kernel: int, expansion: int):
+        super().__init__()
+        self.ff1 = FeedForward(d, expansion)
+        self.mixer = MinGRUMixer(d)
+        self.conv = ConvModule(d, kernel)
+        self.ff2 = FeedForward(d, expansion)
+        self.norm = nn.LayerNorm(d)
+
+    def forward(self, x, state):
+        gru_h, conv_cache = state
+        x = x + 0.5 * self.ff1(x)
+        m, gru_h = self.mixer(x, gru_h)
+        x = x + m
+        c, conv_cache = self.conv(x, conv_cache)
+        x = x + c
+        x = x + 0.5 * self.ff2(x)
+        return self.norm(x), (gru_h, conv_cache)
+
+
+class _Attn(nn.Module):
+    def __init__(self, d: int, heads: int):
+        super().__init__()
+        self.norm = nn.LayerNorm(d)
+        self.mha = nn.MultiheadAttention(d, heads, batch_first=True)
+
+
+class WindowedAttention(nn.Module):
+    """Residual self-attention: the chunk's frames attend to the chunk and up to `left` cached
+    earlier inputs. No positional encoding (the causal convolutions provide position)."""
+
+    def __init__(self, d: int, heads: int, left: int):
+        super().__init__()
+        self.attn = _Attn(d, heads)
+        self.left = left
+
+    def forward(self, x, cache):
+        ctx = torch.cat([cache, x], dim=1)
+        q, kv = self.attn.norm(x), self.attn.norm(ctx)
+        y, _ = self.attn.mha(q, kv, kv, need_weights=False)
+        return x + y, ctx[:, -self.left:]
+
+
+class Encoder(nn.Module):
+    def __init__(self, cfg: dict):
+        super().__init__()
+        d = cfg["d_model"]
+        self.blocks = nn.ModuleList([Block(d, cfg["conv_kernel"], cfg["ffn_expansion"])
+                                     for _ in range(cfg["n_blocks"])])
+        self.attn = nn.ModuleDict({str(p): WindowedAttention(d, cfg["attn_heads"], cfg["attn_left_context"])
+                                   for p in cfg["attn_after_blocks"]})
+
+    def initial_state(self, device):
+        d, k = self.blocks[0].norm.normalized_shape[0], self.blocks[0].conv.dw.kernel_size[0]
+        return {"blocks": [(torch.zeros(1, d, device=device), torch.zeros(1, d, k - 1, device=device))
+                           for _ in self.blocks],
+                "attn": {p: torch.zeros(1, 0, d, device=device) for p in self.attn}}
+
+    def forward(self, x, state):
+        new_blocks, new_attn = [], {}
+        for i, blk in enumerate(self.blocks):
+            x, st = blk(x, state["blocks"][i])
+            new_blocks.append(st)
+            p = str(i + 1)
+            if p in self.attn:
+                x, new_attn[p] = self.attn[p](x, state["attn"][p])
+        return x, {"blocks": new_blocks, "attn": new_attn}
+
+
+# --------------------------------------------------------------------------- front end
+class LogMel(nn.Module):
+    def __init__(self, n_mels=80, n_fft=512, win=400, hop=160):
+        super().__init__()
+        self.n_fft, self.win, self.hop = n_fft, win, hop
+        self.register_buffer("fb", torch.zeros(n_mels, n_fft // 2 + 1))     # loaded from the checkpoint
+        self.register_buffer("window", torch.hann_window(win))
+
+    def forward(self, wav):
+        """[B, samples] -> [B, frames, n_mels], frames = 1 + samples // hop."""
+        spec = torch.stft(wav, self.n_fft, self.hop, win_length=self.win, window=self.window,
+                          center=True, return_complex=True)
+        mel = torch.einsum("mf,bft->bmt", self.fb, spec.real ** 2 + spec.imag ** 2)
+        return torch.log(mel + 1e-6).transpose(1, 2)
+
+
+class Subsampler(nn.Module):
+    """Two Conv2d (3x3, stride 2) over (time, mel) -> 40 ms frames."""
+
+    def __init__(self, n_mels: int, d: int):
+        super().__init__()
+        self.conv = nn.Sequential(nn.Conv2d(1, d, 3, stride=2, padding=1), nn.ReLU(),
+                                  nn.Conv2d(d, d, 3, stride=2, padding=1), nn.ReLU())
+        f_out = ((n_mels - 1) // 2 + 1 - 1) // 2 + 1
+        self.out = nn.Linear(d * f_out, d)
+
+    def forward(self, mel):
+        x = self.conv(mel.unsqueeze(1))                        # [B, d, T/4, F/4]
+        b, c, t, f = x.shape
+        return self.out(x.transpose(1, 2).reshape(b, t, c * f))
+
+
+class FrontEnd(nn.Module):
+    def __init__(self, n_mels: int, d: int):
+        super().__init__()
+        self.logmel = LogMel(n_mels)
+        self.subsample = Subsampler(n_mels, d)
+
+
+# --------------------------------------------------------------------------- transducer
+class PredictionNetwork(nn.Module):
+    def __init__(self, vocab_size: int, hidden: int):
+        super().__init__()
+        self.embed = nn.Embedding(vocab_size, hidden)
+        self.rnn = MinGRU(hidden)
+
+    def step(self, token: int, state):
+        return self.rnn.step(self.embed(torch.tensor([token], device=state.device)), state)
+
+
+class Joiner(nn.Module):
+    def __init__(self, enc_dim: int, pred_dim: int, joint_dim: int, vocab_size: int):
+        super().__init__()
+        self.enc_proj = nn.Linear(enc_dim, joint_dim)
+        self.pred_proj = nn.Linear(pred_dim, joint_dim)
+        self.out = nn.Linear(joint_dim, vocab_size)
+
+
+@dataclass(frozen=True)
+class Word:
+    text: str
+    start: float      # seconds, time of the first character's encoder frame
+    end: float        # seconds, time of the last character's encoder frame
+
+
+@dataclass
+class Transcript:
+    words: list[Word]
+
+    @property
+    def text(self) -> str:
+        return " ".join(w.text for w in self.words)
+
+
+class Mateli(nn.Module):
+    def __init__(self, cfg: dict):
+        super().__init__()
+        self.cfg = cfg
+        d = cfg["d_model"]
+        self.frontend = FrontEnd(cfg["n_mels"], d)
+        self.encoder = Encoder(cfg)
+        self.prednet = PredictionNetwork(len(cfg["vocab"]), cfg["pred_hidden"])
+        self.joiner = Joiner(d, cfg["pred_hidden"], cfg["joint_dim"], len(cfg["vocab"]))
+
+    @classmethod
+    def from_pretrained(cls, model_dir=".", device: str | None = None) -> "Mateli":
+        from safetensors.torch import load_file
+        d = Path(model_dir)
+        model = cls(json.loads((d / "config.json").read_text(encoding="utf-8")))
+        model.load_state_dict(load_file(str(d / "model.safetensors")))
+        device = device or ("cuda" if torch.cuda.is_available() else "cpu")
+        return model.to(device).eval()
+
+    @property
+    def device(self):
+        return self.joiner.out.weight.device
+
+    @torch.inference_mode()
+    def transcribe(self, samples) -> Transcript:
+        """float32 mono 16 kHz samples in [-1, 1] -> words with times."""
+        cfg = self.cfg
+        wav = torch.as_tensor(samples, dtype=torch.float32, device=self.device)
+        if wav.ndim != 1:
+            raise ValueError(f"expected mono samples of shape [n], got {tuple(wav.shape)}; mix channels down first")
+        wav = wav[None]
+        if wav.shape[1] <= self.frontend.logmel.n_fft // 2:
+            return Transcript([])
+        mel = self.frontend.logmel(wav)                         # [1, M, 80]
+        n_chunk, ctx = cfg["chunk_mel_frames"], 4               # 4 mel frames = one 40 ms frame (total stride 4)
+        state = self.encoder.initial_state(self.device)
+        blank, space = cfg["blank_id"], cfg["space_id"]
+        pred, h = self.prednet.step(blank, torch.zeros(1, cfg["pred_hidden"], device=self.device))
+        pred_proj = self.joiner.pred_proj(pred)
+        words, chars, t0, t1, frame = [], "", 0, 0, 0
+        fs = cfg["frame_seconds"]
+        for start in range(0, mel.shape[1], n_chunk):
+            # the subsampler sees the chunk plus 4 earlier mel frames; the first output frame,
+            # which only exists because of that context, is dropped
+            lo = max(0, start - ctx)
+            x = self.frontend.subsample(mel[:, lo:start + n_chunk])
+            x = x[:, 1:] if lo < start else x
+            enc, state = self.encoder(x, state)
+            enc_proj = self.joiner.enc_proj(enc)[0]
+            for e in enc_proj:
+                for _ in range(cfg["max_symbols_per_frame"]):
+                    k = int(self.joiner.out(torch.tanh(e + pred_proj[0])).argmax())
+                    if k == blank:
+                        break
+                    if k == space:
+                        if chars:
+                            words.append(Word(chars, round((t0 + 1) * fs, 2), round((t1 + 1) * fs, 2)))
+                        chars = ""
+                    else:
+                        if not chars:
+                            t0 = frame
+                        chars, t1 = chars + cfg["vocab"][k], frame
+                    pred, h = self.prednet.step(k, h)
+                    pred_proj = self.joiner.pred_proj(pred)
+                frame += 1
+        if chars:
+            words.append(Word(chars, round((t0 + 1) * fs, 2), round((t1 + 1) * fs, 2)))
+        return Transcript(words)
+
+
+def load_audio(path, sample_rate: int = 16000):
+    import numpy as np
+    import soundfile as sf
+    x, sr = sf.read(str(path), dtype="float32", always_2d=True)
+    x = x.mean(axis=1)
+    if sr != sample_rate:
+        from scipy.signal import resample_poly
+        g = math.gcd(sr, sample_rate)
+        x = resample_poly(x, sample_rate // g, sr // g).astype(np.float32)
+    return x
+
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__.split("\n")[0])
+    ap.add_argument("audio")
+    ap.add_argument("--model-dir", default=str(Path(__file__).resolve().parent))
+    ap.add_argument("--device", default=None)
+    ap.add_argument("--words", action="store_true", help="print words with start and end times")
+    a = ap.parse_args()
+    asr = Mateli.from_pretrained(a.model_dir, a.device)
+    t = asr.transcribe(load_audio(a.audio, asr.cfg["sample_rate"]))
+    if a.words:
+        for w in t.words:
+            print(f"{w.start:7.2f} {w.end:7.2f}  {w.text}")
+    else:
+        print(t.text)
+
+
+if __name__ == "__main__":
+    main()

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9f9c1a41a645b41ea24fa54ef9660c1e97813bda21c8622f77767551dac195e8
+size 238443016