whatmodel.sh / stt / Omnilingual ASR LLM 7B v2

Omnilingual ASR LLM 7B v2

+ contribute
Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Omnilingual ASR LLM 7B v2 is an open-weights speech-to-text model from Meta, released December 2025. Its published size is 7.8B parameters. Licence: apache-2.0.

Catalogue identity notes

Omnilingual ASR model card 'omniASR_LLM_7B_v2' (checkpoint https://dl.fbaipublicfiles.com/mms/omniASR-LLM-7B-v2.pt, listed in src/omnilingual_asr/cards/models/rc_models_v2.yaml @ 81f51e2; 'December 2025 Update'). Not on Hugging Face; the CSV entry 'facebook/omniASR-LLM-7B-v2' is the leaderboard's label. Distinct from the v1 checkpoint of the same size. Params: README table. Needs libsndfile (README).

57.6±2.5#50 of 162 speech-to-text modelsfrom 2 benchmark results across 1 source* thinly tested, treat the rank as provisional

Score within the Open ASR English short-form WER (March 2026) comparison set only.

$ how it compares

The speech-to-text models ranked either side of Omnilingual ASR LLM 7B v2, out of 176 with a score, then more from Meta.

#modelscorewhereRAMreleased
57Phi-4-multimodal-instruct Microsoft58±2local~14GB2025-02
58Omnilingual ASR LLM 7B Meta58±2local~19GB2025-11
59Omnilingual ASR LLM 7B v2 Meta58±3local~19GB2025-12
60Distil-Whisper distil-large-v3.5 Hugging Face (Distil-Whisper)58±1local~2.3GB2025-03
61Whisper large-v3-turbo OpenAI57±1api $0.0005/audio min +local~2.4GB2024-10
more from Meta
1Muse Voice Transcribe 1.0 Meta71±3api $0.0030/audio min2026-09
62Omnilingual ASR LLM 3B Meta57±2local~11GB2025-11
75Omnilingual ASR LLM 1B v2 Meta56±3local~6GB2025-12

see all speech-to-text models · find the ones that fit your machine

$ what it's good at

multilingual transcription▂··▃▄█▆▅▄▂·▂ #9 of 36 * · 75% of its score · 1 test

Beats Cohere Transcribe Arabic (07-2026), Nova-3, Qwen3-ASR-1.7B here, though they rank higher overall. Trails SeamlessM4T v2 Large, Whisper large-v3 here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
Open ASR multilingual WER4.4% ✎▂··▂▃▂█▃█▂▁▁ beats 24/29avg 8.9%▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 75%typical
other tests1 not in a task · 22% of its score
testresultvs the others in its rundifficultycountswhy
Open ASR English short-form WER (March 2026)8.1% ✎▁▁▁▃··▁▁▂██▆ beats 33/86avg 8.2%▇▇▇▇▇▇▇▇▇ 22%blunt

$ where to run it

Other ways to run it

  1. Run it as published · needs ~19.22 GB · from github.com/facebookresearch (official)
    uv run --python 3.12 --with omnilingual-asr python -c "from omnilingual_asr.models.inference.pipeline import ASRInferencePipeline as P; print(P(model_card='omniASR_LLM_7B_v2').transcribe(['audio.wav'], batch_size=1))"

Official weights and documentation: github.com/facebookresearch ↗

$ how the score was worked out · Open ASR Leaderboard 98%

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

sourceresultsshare of the scorelicence
Open ASR Leaderboard2▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 98%Apache-2.0
prior · pull toward the middle▇ 2%
weighted average= 57.6

$ improve this page

Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated · sources and licences · json