whatmodel.sh / stt / GPT-4o Transcribe

GPT-4o Transcribe

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

GPT-4o Transcribe is a closed-weights speech-to-text model from OpenAI, released March 2025. Licence: proprietary.

53.2±1.2#83 of 162 speech-to-text modelsfrom 29 benchmark results across 11 sources

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 GPT-4o Transcribe, out of 176 with a score, then more from OpenAI.

#modelscorewhereRAMreleased
92Omnilingual ASR CTC 3B Meta54±2local~7.9GB2025-11
93MOSS-Transcribe-Diarize 0.9B OpenMOSS54±4local~1GB2026-07
94GPT-4o Transcribe OpenAI53±1api $0.0060/audio min2025-03
95LiteWhisper large-v3-turbo-acc Efficient-Speech (LiteASR)53±4local~2.1GB2025-02
96Voxtral Mini 4B Realtime (2602) Mistral53±2api $0.0060/audio min +local~3.2GB2026-02
more from OpenAI
13GPT Transcribe OpenAI64±4api $0.0045/audio min2026-07
48GPT-Realtime-Whisper OpenAI59±3api $0.017/audio min2026-05
61Whisper large-v3-turbo OpenAI57±1api $0.0005/audio min +local~2.4GB2024-10

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

$ what it's good at

other tests29 not in a task · 99% of its score
testresultvs the others in its rundifficultycountswhy
Pipecat STT Benchmark · pooled semantic WER (1000 smart-turn samples, Sonnet 5.5 judge)2.9% ✎▃·▃▅·▅▅·▃▅██ beats 3/16avg 4.8%▇▇▇▇▇ 14%typical
AfriSpeech-MultiBench · 7-region African-accent mean WER38.9% ✎▂····▃···▂▆█ beats 1/11avg 46%▇▇▇▇ 10%typical
AfriSpeech-MultiBench · Parliament (legal domain) WER64.4% ✎▂▃····▂·▅·▃█ beats 1/12avg 52%▇▇▇ 8.7%sharp
Bangla STT benchmark · BanSpeech 1,001-clip batch CER28.7% ✎█·▄··▄··▄·▄▄ beats 1/6avg 37%▇▇▇ 6.9%sharp
Mega-ASR adverse-condition suite · mean WER (CHiME-4, VOiCES, NOIZEUS)17.3% ✎▄····█▄▄···▄ beats 0/5avg 19%▇▇ 6.0%typical
Qwen-Audio-3.0-ASR report · Chinese 4-set CER mean (own API runs)6.6% ✎▄········▄·█ beats 0/3avg 6.7%▇▇ 4.7%sharp
Multilingual LibriSpeech English test WER5.1% ✎█······█·█·█ beats 1/3avg 7.7%▇ 3.4%typical
Qwen-Audio-3.0-ASR report · English 3-set WER mean (own API runs)3.7% ✎█···█··█···█ beats 0/3avg 5.6%▇ 3.3%typical
21 more tests
testresultvs the others in its rundifficultycountswhy
WenetSpeech test_net error rate15.3% ✎█···█·····██ beats 0/3avg 13%▇ 3.2%typical
AISHELL-2 test error rate4.2% ✎█··█·····█·█ beats 1/3avg 5.7%▇ 3.2%typical
verli STT benchmark · LibriSpeech test-clean WER1.5% ✎▃·····▃▅█▃▅▅ beats 11/11avg 3.4%▇ 3.2%typical
GigaSpeechBench · 7 Arabic dialect WER mean50.5% ✎█··█·██··█·█ beats 1/5avg 79%▇ 3.1%typical
LibriSpeech test-other WER3.8% ✎▁▁▁▂▅▃▅▄█▂▂▁ beats 43/54avg 5.5%▇ 2.8%typical
GigaSpeechBench · 12 English domains entity B-WER mean24.8% ✎▄·······▄·█▄ beats 0/4avg 28%▇ 2.7%typical
Perle code-switching benchmark · 4-pair mean WER38.6% ✎█··········▄ beats 1/2avg 55%▇ 2.5%small run
GigaSpeechBench · 6 English accents WER mean39.9% ✎▄······▄··▄█ beats 0/4avg 36%▇ 2.2%typical
Common Voice English WER (version unspecified)9.1% ✎█··▄·······▄ beats 2/3avg 13%▇ 2.2%typical
Common Voice Mandarin error rate (version unspecified)6.3% ✎█·······██·█ beats 2/3avg 11%▇ 2.2%typical
GigaSpeech test WER25.5% ✎▃····▁··▁▄█▂ beats 0/17avg 17%▇ 2.1%typical
Danish ASR Leaderboard · mean WER over 5 Danish test sets (RyeAI)26.1% ✎▂▂···▃▂▅█▃▂▂ beats 9/16avg 32%▇ 2.1%blunt
WenetSpeech test_meeting error rate32.3% ✎█··█······██ beats 0/3avg 21%▇ 2.0%sharp · noisy
koedesk FLEURS first-200 · 9-language headline error mean4.0% ✎▄·····▄··▄█▄ beats 3/5avg 7.9%▇ 1.8%typical
VoxPopuli English test WER10.3% ✎█▄▄··████▄██ beats 0/16avg 13%▇ 1.7%typical
verli STT benchmark · SPGISpeech 1,200-clip sample WER9.5% ✎▃·····▄▂█▂▂▄ beats 0/11avg 4.6%▇ 1.5%typical
verli STT benchmark · Earnings-22 WER (2,741 utt.)18.6% ✎▃·▃▅··▃▃▅▅█▃ beats 0/13avg 16%▇ 1.4%blunt
FLEURS Mandarin test error rate2.4% ✎▄·▂▄▂▄██▂▂·▆ beats 19/20avg 4.3%▇ 1.2%typical
LibriSpeech test-clean WER1.4% ✎▁·▁·▁▁▁▁▃▇█▆ beats 94/100avg 2.1%▇ 0.8%blunt
FLEURS English test WER2.4% ✎▁··▁▁▁▁▂▃█▂▃ beats 64/64avg 4.6%▇ 0.6%blunt
TED-LIUM test WER (short-form)7.7% ✎▁·▁▇█▃·▃▁▃█▄ beats 15/27avg 6.2%▇ 0.1%blunt

$ where to run it

Through an API

api · $0.0060/audio min · developers.openai.com ↗ · price sourceProvider-published estimated cost per audio minute; underlying billing uses audio/text tokens. Exact endpoint gpt-4o-transcribe. Product: Audio transcription. API model: gpt-4o-transcribe. Settings: endpoint=/v1/audio/transcriptions, streaming=no. Price basis: provider-published estimate. Currency: USD.

$ how the score was worked out · Qwen/Qwen3-ASR-1.7B model card 26%, AfriSpeech-MultiBench (arXiv:… 19%, Pipecat STT Benchmark · seman… 14% and 8 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

sourceresultsshare of the scorelicence
Qwen/Qwen3-ASR-1.7B model card13▇▇▇▇▇▇▇▇ 26%Apache-2.0
AfriSpeech-MultiBench (arXiv:2511.14255v1) · African-accented English ASR benchmark suite2▇▇▇▇▇▇ 19%CC BY 4.0
Pipecat STT Benchmark · semantic WER results table (pipecat-ai/stt-benchmark README)1▇▇▇▇ 14%BSD-2-Clause
GigaSpeechBench: A Real-World Multilingual Speech-to-Text Benchmark3▇▇ 8%CC BY 4.0
Qwen-Audio-3.0-ASR Technical Report (arXiv:2609.07549v2)2▇▇ 8%CC BY 4.0
Bangla Speech-to-Text Benchmark v1.0 (Riverborn)1▇▇ 7%MIT
verliapp/stt-benchmark · independent STT accuracy benchmark3▇▇ 6%MIT
Mega-ASR (arXiv:2605.19833v1) · adverse-condition ASR baselines (CHiME-4, VOiCES, NOIZEUS)1▇▇ 6%CC BY 4.0
Benchmarking Commercial ASR Systems on Code-Switching Speech (arXiv:2605.19069v3)1▇ 3%CC BY 4.0
Danish Speech-to-Text Leaderboard (RyeAI) · results dataset1▇ 2%MIT
koedesk STT Bench v1 (Track A, FLEURS first-200)1▇ 2%CC BY 4.0
prior · pull toward the middle▇ 1%
weighted average= 53.2

$ 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