TranscribePsy
Collection
Compact multilingual Automatic Speech Recognition (ASR) model that transcribes speech to text, and purpose-built to run on the phone in your pocket. • 4 items • Updated • 1
TranscribePsy-250M-EU-sub
TranscribePsy-250M-EU-sub is a compact (254M-parameter) speech recognition model covering the eight European languages selected by Multilingual LibriSpeech (MLS), distilled from Parakeet-TDT-0.6B-v3 and built for on-device and edge deployment. Concentrating the same parameter budget on fewer languages makes it the most accurate model in the TranscribePsy family on the languages it covers.
- Lowest error in the family: 7.50 on the Open ASR Leaderboard and 5.87 on Multilingual LibriSpeech, ahead of the 25-language TranscribePsy-250M-EU (8.02 / 6.58).
- Beats models three to four times larger: clearly ahead of Qwen3-ASR-0.6B (782M), omniASR-CTC-1B-v2 (976M) and Whisper-Medium (764M) on Multilingual LibriSpeech.
- Ships for the phone: available as
GGUF for our
qvac-fabric-speech.cppengine and Core ML for the Apple Neural Engine.
Supported languages (8): Dutch (nl), English (en), French (fr), German (de), Italian (it), Polish (pl), Portuguese (pt), Spanish (es).
Which model should I use?
| Model | Languages | Format | Use it when |
|---|---|---|---|
| TranscribePsy-250M-EU | 25 | PyTorch | You need broad European language coverage. |
| TranscribePsy-250M-EU-sub (this card) | 8 | PyTorch | You only need the eight Multilingual LibriSpeech languages and want the lowest error. |
| TranscribePsy-250M-EU-sub-gguf | 8 | GGUF (Q8_0) | You are deploying on-device through our qvac-fabric-speech.cpp engine. |
| TranscribePsy-250M-EU-sub-coreml | 8 | Core ML | You are deploying on the Apple Neural Engine. |
Where TranscribePsy-250M-EU-sub leads
Lower is better throughout.
| Benchmark suite | TranscribePsy-250M-EU-sub | TranscribePsy-250M-EU | Whisper-small (242M) | omniASR-CTC-300M-v2 (326M) | OWSM v4 small (370M) |
|---|---|---|---|---|---|
| Open ASR Leaderboard | 7.50 | 8.02 | 12.34 | 15.31 | 9.72 |
| Multilingual LibriSpeech | 5.87 | 6.58 | 11.51 | 11.89 | 11.84 |
| Common Voice v25 | n/a * | 21.56 | 38.17 | 32.23 | 37.74 |
* The Common Voice suite spans languages outside this model's eight, so no comparable mean is reported. The other means are taken over all 24 languages of the test set. Use TranscribePsy-250M-EU for broader coverage.
Against models three to four times larger
| Metric | TranscribePsy-250M-EU-sub | Qwen3-ASR-0.6B (782M) | omniASR-CTC-1B-v2 (976M) | Whisper-Medium (764M) |
|---|---|---|---|---|
| Parameters | 254M | 782M | 976M | 764M |
| Open ASR Leaderboard | 7.50 | 7.40 | 9.68 | 8.79 |
| Multilingual LibriSpeech | 5.87 | 12.60 | 7.71 | 7.76 |
| All on-device figures below are measured on phones with 8-bit quantized builds; TranscribePsy-250M-EU-sub runs as its GGUF build. | ||||
| Model size (MB) | 305 | 854 | 1,056 | 823 |
| Real-time factor | 0.052 | 0.462 | 0.998 | 0.931 |
| Inference latency (ms) | 1,021 | 7,616 | 15,372 | 14,636 |
| Cold-start latency (ms) | 252 | 677 | 898 | 870 |
| Peak memory @ 10s (MiB) | 534 | 1,169 | 1,275 | 1,569 |
Per-dataset results
Every test set behind the three suites above, against the same-size models. Word error rate, lower is better; bold marks the best score in each row.
Open ASR Leaderboard: 7 English short-form and 9 multilingual test sets
The reported Open ASR Leaderboard score is the mean of the English short-form mean and the multilingual mean, which is itself the mean of the FLEURS mean and the MLS mean.
| Test set | TranscribePsy-250M-EU-sub | Whisper-small | omniASR-CTC-300M-v2 | OWSM v4 small |
|---|---|---|---|---|
| AMI | 12.37 | 53.35 | 32.51 | 15.47 |
| Earnings-22 | 14.77 | 13.21 | 26.33 | 14.48 |
| GigaSpeech | 11.23 | 10.56 | 22.84 | 11.17 |
| LibriSpeech clean | 2.14 | 2.59 | 4.71 | 2.56 |
| LibriSpeech other | 4.81 | 6.60 | 11.86 | 5.98 |
| SPGISpeech | 8.77 | 3.24 | 14.05 | 2.79 |
| VoxPopuli | 7.42 | 7.85 | 13.28 | 8.65 |
| English short-form mean | 8.79 | 13.91 | 17.94 | 8.73 |
| FLEURS German | 6.58 | 7.61 | 14.53 | 10.08 |
| FLEURS Spanish | 4.76 | 4.67 | 9.56 | 8.21 |
| FLEURS French | 8.24 | 11.52 | 21.23 | 13.16 |
| FLEURS Italian | 3.76 | 16.02 | 9.46 | 6.87 |
| FLEURS Portuguese | 7.32 | 6.27 | 13.63 | 8.76 |
| FLEURS mean | 6.13 | 9.22 | 13.68 | 9.42 |
| MLS Spanish | 4.20 | 6.89 | 7.25 | 7.61 |
| MLS French | 4.07 | 10.93 | 12.95 | 11.05 |
| MLS Italian | 6.58 | 16.02 | 10.73 | 14.49 |
| MLS Portuguese | 10.41 | 15.37 | 15.81 | 14.83 |
| MLS mean | 6.32 | 12.30 | 11.69 | 12.00 |
| Multilingual mean (FLEURS & MLS) | 6.22 | 10.76 | 12.68 | 10.71 |
| Open ASR Leaderboard score | 7.50 | 12.34 | 15.31 | 9.72 |
Multilingual LibriSpeech: 8 languages
| Language | TranscribePsy-250M-EU-sub | Whisper-small | omniASR-CTC-300M-v2 | OWSM v4 small |
|---|---|---|---|---|
| English | 5.68 | 7.66 | 15.10 | 7.96 |
| German | 5.99 | 9.35 | 11.02 | 10.43 |
| Dutch | 5.09 | 16.54 | 12.89 | 11.64 |
| Polish | 4.89 | 9.26 | 9.39 | 16.72 |
| Spanish | 4.21 | 6.90 | 7.25 | 7.61 |
| French | 4.06 | 10.94 | 12.95 | 11.05 |
| Italian | 6.57 | 16.06 | 10.73 | 14.49 |
| Portuguese | 10.46 | 15.37 | 15.81 | 14.83 |
| Mean (8 languages) | 5.87 | 11.51 | 11.89 | 11.84 |
Common Voice v25
Only the eight languages this model covers are shown, so no suite-level mean is reported.
| Language | TranscribePsy-250M-EU-sub | Whisper-small | omniASR-CTC-300M-v2 | OWSM v4 small |
|---|---|---|---|---|
| German | 7.37 | 13.97 | 19.29 | 10.71 |
| English | 12.67 | 17.18 | 31.20 | 14.22 |
| Spanish | 5.99 | 11.63 | 14.23 | 9.34 |
| French | 9.99 | 26.92 | 24.82 | 16.22 |
| Italian | 5.94 | 16.81 | 16.28 | 10.24 |
| Dutch | 4.55 | 12.66 | 16.74 | 12.11 |
| Polish | 13.52 | 20.74 | 25.53 | 24.85 |
| Portuguese | 9.02 | 22.75 | 23.69 | 13.01 |
Open ASR Leaderboard multilingual benchmark (October 2026 version)
Evaluated with the official Open ASR Leaderboard multilingual scripts (NeMo 2.7.2) on its 16 test sets: FLEURS, MCV (CoVoST-2) and MLS across six languages. Averages follow the leaderboard: per language, then across languages. WER, lower is better.
| Language | FLEURS | MCV (CoVoST-2) | MLS | Language average |
|---|---|---|---|---|
| German | 6.58 | 6.75 | – | 6.67 |
| French | 8.23 | 8.31 | 4.07 | 6.87 |
| Italian | 3.76 | 5.37 | 6.58 | 5.24 |
| Spanish | 4.76 | 5.43 | 4.20 | 4.80 |
| Portuguese | 7.32 | – | 10.41 | 8.87 |
| Dutch | 9.30 | 7.21 | 5.11 | 7.21 |
| Average (6 languages) | 6.61 |
Model details
| Property | Value |
|---|---|
| Developed by | Tether AI Research * |
| Parameters | 253.95M |
| Architecture | FastConformer encoder, 24 layers, 640-dim, transducer decoder |
| Teacher | Parakeet-TDT-0.6B-v3 (627M) |
| Initialization | Structured pruning of the teacher |
| Languages | 8 |
| Precision | float32 |
| Model size on disk | 1.02 GB (float32) |
* References to Tether AI Research are references to Tether Data, S.A. de C.V.
Languages
Dutch, English, French, German, Italian, Polish, Portuguese, Spanish.
On-device efficiency
Measured on six phones: Samsung Galaxy S26 Ultra, Samsung Galaxy S25 Ultra, Google Pixel 10,
Google Pixel 9, iPhone 17 Pro Max, and iPhone 15. Every model runs as an 8-bit build under the same test settings, so the comparison reflects the models rather than their inference stacks; TranscribePsy runs through our qvac-fabric-speech.cpp engine. The benchmark set is 160 utterances drawn at random from Multilingual LibriSpeech, 20 per language across eight languages, 40 minutes of audio in total. Every figure is the mean of the best three of five repeated runs, which keeps run-to-run thermal throttling out of the numbers.
| Device | SoC | Inference latency (ms) | Cold start (ms) | Peak memory @ 10s (MiB) |
|---|---|---|---|---|
| Galaxy S26 Ultra | 8 Elite Gen5 | 633 | 213 | 530 |
| Galaxy S25 Ultra | 8 Elite | 771 | 241 | 533 |
| iPhone 17 Pro Max | A19 Pro | 1,059 | 153 | 542 |
| iPhone 15 | A16 | 854 | 221 | 537 |
| Pixel 10 | Tensor G5 | 1,240 | 327 | 530 |
| Pixel 9 | Tensor G4 | 1,570 | 356 | 530 |
TranscribePsy is the fastest model on every one of these devices.
Usage
One set of weights covers all eight languages, so there is nothing to route and no language to declare.
| File | Size | What it is |
|---|---|---|
TranscribePsy-250M-EU-sub.nemo |
1.02 GB | The model, as a NeMo checkpoint |
model.safetensors |
1.02 GB | The same weights in safetensors |
inference.py |
Reference transcription script | |
eval/ |
Decoding and scoring scripts behind the benchmark numbers |
Requirements: nemo_toolkit[asr], torch. Audio must be 16 kHz mono.
import nemo.collections.asr as nemo_asr
model = nemo_asr.models.ASRModel.from_pretrained("qvac/TranscribePsy-250M-EU-sub").eval()
transcripts = model.transcribe(["audio_en.wav", "audio_de.wav", "audio_pt.wav"])
print([t.text for t in transcripts])
Or from the command line, after hf download qvac/TranscribePsy-250M-EU-sub --local-dir TranscribePsy-250M-EU-sub:
python inference.py audio_en.wav audio_de.wav audio_pt.wav
For on-device deployment use the
GGUF build with our qvac-fabric-speech.cpp engine, or
the Core ML build on Apple
silicon.
Intended use
TranscribePsy targets latency- and memory-constrained on-device transcription: voice notes, dictation, meeting capture, accessibility captioning, and voice interfaces that must work offline. Because it is compact, we recommend fine-tuning on your domain when accuracy on specialist vocabulary matters.
Limitations
- Eight languages only. Dutch, English, French, German, Italian, Polish, Portuguese, and Spanish. For wider coverage use TranscribePsy-250M-EU.
- Batch transcription, not streaming. This release transcribes complete utterances.
- Not intended for safety-critical or high-stakes automated decisions.
Acknowledgements
Built on NVIDIA's Parakeet-TDT and FastConformer, and on the NeMo toolkit. Evaluated with the Open ASR Leaderboard protocol and text normalizer.
Citation
A citation for this model is coming soon.
Copyright
We will take appropriate actions in response to notices of copyright infringement. If you believe your work has been used or copied in a manner that infringes upon your intellectual property rights, please email data-apps@tether.io identifying and describing both the copyrighted work and alleged infringing content.
Licensing
This model which was trained as described in the blog post (https://huggingface.co/blog/qvac/transcribepsy-250m) is licensed by Tether Data, S.A. de C.V. under the Apache 2.0 license. As described in the blog post, this model is a version of the parakeet-tdt-0.6b-v3 (https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3) which is licensed under the CC-BY-4.0 license (Creative Commons Attribution 4.0 International).
As described in the blog post, the parakeet-tdt-0.6b-v3 model was used as a teacher model.
The Yodas-granary dataset (https://huggingface.co/datasets/espnet/yodas-granary) is made available under the CC-BY-3.0 (Creative Commons Attribution 3.0 Unported). Yodas-granary is a subset of the Granary dataset (https://huggingface.co/datasets/nvidia/Granary) which is made available under the CC-BY-4.0 license (Creative Commons Attribution 4.0 International). The Voxpopuli dataset (https://github.com/facebookresearch/voxpopuli?tab=License-1-ov-file) is made available under the CC0-1.0 license (Creative Commons Zero v1.0 Universal). The Libri-light dataset (https://github.com/facebookresearch/libri-light) is made available under the MIT License (https://github.com/facebookresearch/libri-light?tab=MIT-1-ov-file). The Multilingual LibriSpeech (MLS) dataset (https://www.openslr.org/94/) is made available under the CC-BY-4.0 license (Creative Commons Attribution 4.0 International). The AMI Corpus dataset (https://groups.inf.ed.ac.uk/ami/corpus/) is made available under the CC-BY-4.0 license (Creative Commons Attribution 4.0 International). The LibriSpeech ASR Corpus (960 hours) dataset (https://www.openslr.org/12) is made available under the CC-BY-4.0 license (Creative Commons Attribution 4.0 International). The National Speech Corpus dataset (https://www.imda.gov.sg/how-we-can-help/national-speech-corpus) is made available under the Singapore Open Data Licence version 1.0 (https://data.gov.sg/open-data-licence). The VCTK Corpus (https://www.kaggle.com/datasets/pratt3000/vctk-corpus) is made available under the ODC-BY v1.0 (Open Data Commons Attribution License). The Europarl-ASR dataset (https://www.mllp.upv.es/git-pub/ggarces/Europarl-ASR) is made available under the CC-BY-4.0 license (Creative Commons Attribution 4.0 International) except that: (a) speech and text data from the European Parliament website (audio, official transcriptions and translations) incorporated into the Europarl-ASR dataset are the exclusive property of the European Union represented by the European Parliament and these data are reused here under the conditions stated European Parliament website's Legal notice (https://www.europarl.europa.eu/legal-notice/en/); (b) text data from the DCEP Digital Corpus of the European Parliament incorporated into the Europarl-ASR dataset are the exclusive property of the European Parliament and remain owned by the European Parliament and these data are reused here under the usage conditions of the DCEP Digital Corpus of the European Parliament (https://joint-research-centre.ec.europa.eu/language-technology-resources/dcep-digital-corpus-european-parliament_en#Usage%20Conditions); and (c) text data from the Europarl v10 corpus (https://www.statmt.org/europarl/) are reused here under the Europarl corpus terms of use. The Common Voice Scripted Speech 25.0 dataset (https://commonvoice.mozilla.org/en) is made available under the CC0-1.0 license (Creative Commons Zero v1.0 Universal).
As described in the blog post, the parakeet-tdt-0.6b-v3 model was pruned and trained using the Yodas-granary, Voxpopuli, Libri-light, Multilingual LibriSpeech (MLS), National Speech Corpus, VCTK Corpus, AMI, LibriSpeech, and Europarl-ASR datasets and the following languages from the Common Voice Scripted Speech 25.0 dataset: bg (Bulgarian), cs (Czech), da (Danish), de (German), el (Greek), en (English), es (Spanish), et (Estonian), fi (Finnish), fr (French), hu (Hungarian), it (Italian), lt (Lithuanian), lv (Latvian), mt (Maltese), nl (Dutch), pl (Polish), pt (Portuguese), ro (Romanian), ru (Russian), sk (Slovak), sl (Slovenian), sv (Swedish), and uk (Ukrainian).
- Downloads last month
- 5