whatmodel.sh / stt / Whisper medium
Whisper medium
↗ share
+ contribute
Whisper medium is an open-weights speech-to-text model from OpenAI, released September 2022. Its published size is 0.769B parameters. Licence: apache-2.0.
Catalogue identity notes
Multilingual medium checkpoint (openai/whisper-medium@abdf7c3). Hugging Face mirror created 2022-09-26 with the Whisper release (arXiv 2212.04356). Multilingual checkpoint (openai-whisper name 'medium'), distinct from the English-only medium.en. Params from the Whisper README size table.
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 Whisper medium, out of 176 with a score, then more from OpenAI.
| # | model | score | where | ||
|---|---|---|---|---|---|
| 140 |  ESPnet Branchformer 18-layer (70 epochs, no LM) | 44±4 | local | ||
| 141 |  Nemotron 3.5 ASR Streaming 0.6B | 44±3 | local | ||
| 142 | Whisper medium | 43±2 | local | ||
| 143 |  Omnilingual ASR CTC 300M v2 | 43±3 | local | ||
| 144 | STT En Conformer-Transducer Small | 43±3 | local | ||
| more from OpenAI | |||||
| 13 | GPT Transcribe | 64±4 | api $0.0045/audio min | ||
| 48 | GPT-Realtime-Whisper | 59±3 | api $0.017/audio min | ||
| 61 | Whisper large-v3-turbo | 57±1 | api $0.0005/audio min +local | ||
see all speech-to-text models · find the ones that fit your machine
$ what it's good at
other tests14 not in a task · 99% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| LibriSpeech test-clean WER +2 | 2.6% ✎ | ▁···▁▁▂·▄▄██ beats 15/56 | 22% |
| LibriSpeech test-other WER | 6.5% ✎ | █▃▃·▃▅·▅···▅ beats 2/11 | 18% |
| Omnilingual ASR paper · FLEURS-81 test mean CER (Whisper-covered languages) | 47.8% ✎ | ▅···▃···▃▃██ beats 1/10 | 9.3% |
| LiteASR paper · ESB 8-set average WER | 14.8% ✎ | ▃·······▂··█ beats 0/6 | 8.6% |
| FLEURS English test WER | 4.6% ✎ | ▁··▁▁▁▁▁▁▃█▁ beats 33/56 | 6.5% |
| Omnilingual ASR paper · Common Voice 22 (76 languages) test mean CER | 87.9% ✎ | ▂▃▂····▂▆··█ beats 1/10 | 6.5% |
| FLEURS Mandarin test error rate | 13.1% ✎ | ▂▂··▂▄▂▃·▆██ beats 6/17 | 6.2% |
| PESTE · Persian FLEURS (fa_ir) corpus CER | 17.0% ✎ | ▄··▄·▄▄▄··██ beats 4/8 | 6.2% |
6 more tests
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Betthupferl (German dialect ASR benchmark) · Standard German audio → Standard German transcript WER | 18.3% ✎ | ▄·█·▄▄▄▄·█▄▄ beats 5/10 | 3.8% |
| Open Universal Arabic ASR Leaderboard · 6-test-set average WER | 45.6% ✎ | ▃·▂▃▅█▅▃▂▅▂▂ beats 3/22 | 3.7% |
| TurkASR-Bench (7 Turkish sources, 5,200 utterances) · corpus WER | 16.1% ✎ | ▃··▃▃··▅··█▅ beats 5/9 | 3.0% |
| AfriSpeech-200 test (108 African accents) · WER | 33.2% ✎ | ███····█████ beats 6/7 | 2.0% |
| Betthupferl (German dialect ASR benchmark) · dialect audio → Standard German transcript WER | 43.1% ✎ | ▃█····▃▅▃·▅▃ beats 5/10 | 1.7% |
| TED-LIUM test WER (short-form) | 12.0% ✎ | ▂····▂▂▂█▃▃▂ beats 12/13 | 0.9% |
$ where to run it
Other ways to run it
version
- Run it as published · needs ~2.3 GB · from github.com/openai (official)
uvx --from openai-whisper whisper audio.mp3 --model medium
Official weights and documentation: huggingface.co/openai ↗
$ how the score was worked out · transcribe.cpp model catalog… 29%, ESPnet LibriSpeech ASR recipe… 23%, Omnilingual ASR: Open-Source… 16% and 7 more
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
$ improve this page
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