Search public pages, research tools, and SAVRN solutions.
SAVRN Model Hub · Models by Task
78 open-weight speech recognition models in the SAVRN Model Hub, with Jonatas Grosman, Systran and Handy publishing the most.
On the speech recognition shelf the download leaders are not the big labs. Of the 78 models we track, 11 come from Jonatas Grosman, one publisher fine-tuning the wav2vec2-large-xlsr-53 base per language on Common Voice 6.1, and the Japanese version pulls 17.8 million downloads a month, ahead of anything from OpenAI. Systran and Handy hold 6 each, OpenAI and Efwkjn 5. Argmax's whisperkit-coreml sits second at 11.2 million and targets Apple Silicon, not a GPU rack. Pyannote's two gated diarization pipelines, 8.2 and 5.2 million, do a different job: 16 kHz mono in, who spoke when out.
Hardware is the short conversation. Parameters run from 22 million for Efwkjn's whisper-ja-22M to 8.7 billion for Microsoft's VibeVoice-ASR, which reads a 60-minute recording in one pass and labels speaker, timestamp and text. OpenAI's whisper-large-v3-turbo, 809 million parameters, needs 1.9 GB at 16-bit and 0.5 GB at 4-bit; whisper-large-v3 at 1.5 billion needs 3.7 GB; even VibeVoice-ASR fits in 20.8 GB. All of them run on the cheapest Index slot, one MI300X at $1.85 an hour, so concurrent audio streams size the machine, not footprint.
Licenses help: 36 Apache 2.0, 12 MIT, 8 CC BY 4.0. The catches are 11 with no license stated, which we would not deploy without the publisher's word, 3 under CC BY-NC 4.0 with no commercial use, and 5 marked other. Before committing, confirm your pipeline delivers 16 kHz audio, which the wav2vec2 fine-tunes and Pyannote pipelines both require, and decide whether you want a separate diarization model or VibeVoice-ASR doing both jobs in one.
| Model | Publisher | Parameters | License | Monthly downloads | Cheapest GPUs at 16-bit |
|---|---|---|---|---|---|
| wav2vec2-large-xlsr-53-japanese | Jonatas Grosman | — | apache-2.0 | 17.8M | — |
| whisperkit-coreml | Argmax | — | mit | 11.2M | — |
| speaker-diarization-3.1 | Pyannote | — | mit | 8.2M | — |
| whisper-large-v3-turbo | OpenAI | 809M | mit | 6.8M | 1x MI300X, $1.85/hr |
| wav2vec2-large-xlsr-53-portuguese | Jonatas Grosman | — | apache-2.0 | 6M | — |
| speaker-diarization-community-1 | Pyannote | — | cc-by-4.0 | 5.2M | — |
| whisper-large-v3 | OpenAI | 1.5B | apache-2.0 | 4.8M | 1x MI300X, $1.85/hr |
| wav2vec2-large-xlsr-53-russian | Jonatas Grosman | — | apache-2.0 | 4.2M | — |
| wav2vec2-large-xlsr-53-polish | Jonatas Grosman | — | apache-2.0 | 3.7M | — |
| wav2vec2-large-xlsr-53-dutch | Jonatas Grosman | — | apache-2.0 | 3.1M | — |
| License | Models | Commercial use |
|---|---|---|
| apache-2.0 | 36 | Yes |
| mit | 12 | Yes |
| not stated | 11 | Not stated |
| cc-by-4.0 | 8 | Yes |
| other | 5 | Read the license |
| cc-by-nc-4.0 | 3 | Not without separate permission |
| cc0-1.0 | 1 | Read the license |
| cc-by-sa-4.0 | 1 | Yes |
| Publisher | Models |
|---|---|
| Jonatas Grosman | 11 |
| Systran | 6 |
| Handy | 6 |
| OpenAI | 5 |
| Efwkjn | 5 |
| Pyannote | 3 |
openai/whisper-tiny.en with ONNX weights to be compatible with Transformers.js. If you haven't already, you can install the Transformers.js JavaScript library from NPM using: Note: Having a separate repo for ONNX weights is intended to be a temporary solution until WebML gains more traction. If you would like to make your models web-ready, we recommend converting to ONNX using Optimum and structuring your repo like this one (with ONNX weights located in a subfolder named onnx).
This repository contains the conversion of openai/whisper-medium to the CTranslate2 model format. This model can be used in CTranslate2 or projects based on CTranslate2 such as faster-whisper. The original model was converted with the following command: Note that the model weights are saved in FP16. This type can be changed when the model is loaded using the computetype option in CTranslate2. For more information about the original model, see its model card.
This repository contains the conversion of openai/whisper-small.en to the CTranslate2 model format. This model can be used in CTranslate2 or projects based on CTranslate2 such as faster-whisper. The original model was converted with the following command: Note that the model weights are saved in FP16. This type can be changed when the model is loaded using the computetype option in CTranslate2. For more information about the original model, see its model card.
GGUF conversions of nvidia/canary-180m-flash for use with transcribe.cpp. Ported from upstream commit pinned 2026-05-08. Validated against the NeMo reference at transcribe.cpp commit Offline multilingual speech-to-text and translation. A 182M-parameter multitask AED with a 17-layer FastConformer encoder and a 4-layer Transformer decoder. Supports automatic speech recognition in English, German, Spanish, and French, and bidirectional EN↔{DE, ES, FR} translation. Takes a 16 kHz mono WAV and produces a transcript. Not a streaming model; word/segment timestamps are upstream-experimental and not exposed in the v1 port. WER on the full LibriSpeech test-clean split (2,620 utterances), batch size…
GGUF conversions of nvidia/canary-1b-v2 for use with transcribe.cpp. Ported from upstream commit pinned 2026-05-08. Validated against the NeMo reference at transcribe.cpp commit Offline multilingual speech-to-text and translation across 25 European languages. A 978M-parameter multitask AED with a 32-layer FastConformer encoder and an 8-layer Transformer decoder. Supports automatic speech recognition for any of the 25 supported languages, plus translation between supported language pairs (per the upstream model card). Takes a 16 kHz mono WAV and produces a transcript. Not a streaming model; word and segment timestamps from the upstream model are not exposed in the v1 port. WER on the full…
GGUF conversions of nvidia/canary-1b-flash for use with transcribe.cpp. Ported from upstream commit pinned 2026-05-08. Validated against the NeMo reference at transcribe.cpp commit Offline multilingual speech-to-text and translation. An 883M-parameter multitask AED with a 32-layer FastConformer encoder and a 4-layer Transformer decoder. Supports automatic speech recognition in English, German, Spanish, and French, and bidirectional EN↔{DE, ES, FR} translation. Takes a 16 kHz mono WAV and produces a transcript. Not a streaming model; word/segment timestamps are upstream-experimental and not exposed in the v1 port. WER on the full LibriSpeech test-clean split (2,620 utterances), batch size 1…
For usage instructions follow openai/whisper-large-v3-turbo. Note for faster-whisper vocab changes make model.ismultilingual and suppresstokens wrong. Please adjust the code as required if you want to use this with faster-whisper. Turbo finetune with japanese tokenizer. Full finetune trained 2^19 steps, batch size 64. Smaller vocab with ~1.6x bytes/token allows faster speed with 4 layers vs 2 layer distil (10% larger decoder). Benchmarks. Short form slightly behind v0.2 (trained less?) but long form much better. Also trained for lyrics but untested. Research supported with Cloud TPUs from Google's TPU Research Cloud (TRC)
For usage instructions follow openai/whisper-large-v3. Large-v3 finetune trained as a baseline with smaller checkpoints in progress. Expecting worse long form and equal short form. Benchmarks. Has occasional repetition issue compared to previous models but achieves competitive/SOTA CER across all tested sets.
Whisper finetune for Japanese focused on general/anime domains. For usage instructions follow openai/whisper-large-v3-turbo. Due to vocab changes ctranslate2>=4.7.1 required for faster-whisper. For inference engines with hardcoded vocab, the token embedding can be padded. Finetuned from turbo with pruned vocab, indices can be found in mapping.txt. Trained decoder only for 2^20 steps, batch size 64. Using a 45000 hour corpus (largest source is 17000 of filtered reazonspeech-all) with custom mixing ratio and augmentation to maintain long form performance and timestamps. Benchmarks. Competitive/SOTA on test sets, slightly better than 1.5B on short form, worse on long form. Also trained for…
Whisper finetune for Japanese focused on general/anime domains. For usage instructions follow openai/whisper-large-v3-turbo. Due to vocab changes ctranslate2>=4.7.1 required for faster-whisper. For inference engines with hardcoded vocab, the token embedding can be padded. Finetuned from base with pruned vocab and encoder conv adaption, indices can be found in mapping.txt. Trained decoder only for 2^20 steps, batch size 64. Using a 45000 hour corpus (largest source is 17000 of filtered reazonspeech-all) with custom mixing ratio and augmentation to maintain long form performance and timestamps. Benchmarks. Competitive for size on test sets, particually good on JSUT-book. Also trained for…
IPA phoneme recognition for Kölsch (Ripuarian German, Cologne). Fine-tuned from facebook/wav2vec2-xls-r-300m with a CTC head over a 48-symbol IPA inventory. This is also the forced-alignment model. Its frame-level CTC posteriors drive torchaudio.functional.forcedalign to produce Praat TextGrids with word and phone tiers — see notebook 9 in the repository. Trained during the CIF Tandem Fellowship at IfL-Phonetik, University of Cologne, on Alles Kölsch (Bhatt & Lindlar 1998) — 4,670 utterances, 4.5 hours of spontaneous Cologne dialect speech from 105 speakers aged 10–88 across 49 neighbourhoods. Kölsch is Ripuarian German. It has no public speech dataset and no standardised spelling: 94.1 %…
Whisper finetune for Japanese focused on general/anime domains. For usage instructions follow openai/whisper-large-v3-turbo. Due to vocab changes ctranslate2>=4.7.1 required for faster-whisper. For inference engines with hardcoded vocab, the token embedding can be padded. Finetuned from tiny with pruned vocab and encoder conv adaption, indices can be found in mapping.txt. Trained decoder only for 2^19 steps, batch size 64. Using a 45000 hour corpus (largest source is 17000 of filtered reazonspeech-all) with custom mixing ratio and augmentation to maintain long form performance and timestamps. Benchmarks. CER roughly between OpenAI whisper-base/small, not great but also the smallest model…
An equal-weight average of the epoch 6, 7, 8, 9 and 10 checkpoints of No training was done here, and inference costs exactly what one model costs. Like its ingredient, this model has no reported WER or CER and cannot have one. That run trains on every held-out hour the project has, including the SAPC2 dev split the rest of this family scores against. Averaging its epochs does not create a set to measure on. So this checkpoint rests on a bet rather than a measurement, and it is worth same soup of the same five epochs was worth 0.27 CER points — 6.06% against 6.33% for the best single epoch. That is the whole of the evidence. It is evidence from a different architecture (RNN-T, not TDT) on…
nvidia/parakeet-tdt-0.6b-v3 fine-tuned on all 1,047.7 hours this project holds: SAPC1 train and dev, SAPC2 train, the SAPC2 dev split the rest of this family scores against, 103.1 hours of synthetic dysarthric speech, 79.2 hours recovered by force-aligning and cutting recordings past the 45-second training cap, and 15.5 hours of AtaxiaUK and HeyJay!, which are outside the challenge corpora and make this an unconstrained-track model. This model has no reported WER or CER, and cannot have one. Every held-out hour is in its training data. That was the point: the hyperparameters were settled on the sibling runs that do hold out a dev split, and this run spends that split as training data…
Fine-tuned openai/whisper-large-v3-turbo on Monsoon Hindi ASR data using LoRA (r=32, alpha=64). English capabilities preserved with <10% relative degradation across all benchmarks. LoRA's structural separation (frozen base + low-rank adapters) provides a natural safeguard against catastrophic forgetting.
Quantized MLX weights for beshkenadze/cohere-transcribe-03-2026-mlx-fp16. - model.safetensors - config.json - tokenizer.model - tokenizerconfig.json - preprocessorconfig.json - specialtokensmap.json - keymap.json - conversionsummary.json This checkpoint has been re-validated against the current Swift and Python MLX runtimes. Verified semantic parity on an English fixture: - official CUDA reference path (transformers native Cohere ASR) Fastest and smallest, but introduces a lexical regression on the repo sample (Kaldi → Khaldi). - Generated from the Swift-compatible fp16 checkpoint beshkenadze/cohere-transcribe-03-2026-mlx-fp16. - This repository contains inference artifacts only. Refer to…
Quantized MLX weights for beshkenadze/cohere-transcribe-03-2026-mlx-fp16. - model.safetensors - config.json - tokenizer.model - tokenizerconfig.json - preprocessorconfig.json - specialtokensmap.json - keymap.json - conversionsummary.json This checkpoint has been re-validated against the current Swift and Python MLX runtimes. Verified semantic parity on an English fixture: - official CUDA reference path (transformers native Cohere ASR) Matches fp16 on the repo sample while reducing memory substantially. - Generated from the Swift-compatible fp16 checkpoint beshkenadze/cohere-transcribe-03-2026-mlx-fp16. - This repository contains inference artifacts only. Refer to the upstream Cohere model…
This is the model card of a transformers model that has been pushed on the Hub. Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. Use the code below to get started with the model. Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).
By monthly downloads reported by the Hugging Face Hub: whisper-tiny.en (795.1k); faster-whisper-medium (752k); faster-whisper-small.en (740.3k).