Whisper large-v3

+ contribute

Whisper large-v3 is an open-weights speech-to-text model from OpenAI, released November 2023. Its published size is 1.55B parameters. Licence: apache-2.0.

56.7±0.8#58 of 162 speech-to-text modelsfrom 80 benchmark results across 31 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 Whisper large-v3, out of 176 with a score, then more from OpenAI.

#	model	score	where	RAM	released
65	Parakeet TDT 0.6B v3 NVIDIA	57±1	api $0.0015/audio min +local	~2GB	2025-08
66	Kyutai STT 2.6B (en) Kyutai	57±4	local	~6.7GB	2025-06
67	Whisper large-v3 OpenAI	57±1	api $0.0018/audio min +local	~4.2GB	2023-11
68	GLM-ASR-Nano-2512 Z.ai	57±4	local	~5.9GB	2025-12
69	Parakeet RNNT 0.6B NVIDIA	57±3	local	~3.5GB	2024-01
	more from OpenAI
13	GPT Transcribe OpenAI	64±4	api $0.0045/audio min		2026-07
48	GPT-Realtime-Whisper OpenAI	59±3	api $0.017/audio min		2026-05
61	Whisper large-v3-turbo OpenAI	57±1	api $0.0005/audio min +local	~2.4GB	2024-10

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

$ what it's good at

long-form transcription▁▁▄▁·▁▂█▄▁▁▁ #5 of 27 * · 2.2% of its score · 1 test

Beats Chirp 3, Parakeet TDT 0.6B v2, Parakeet TDT 1.1B here, though they rank higher overall.

test	result	vs the others in its run	difficulty	counts	why
Open ASR English long-form WER	8.7% ✎	▁▁▅▁·▃▅█▆▁▂▂ beats 23/30	avg 12%	▇ 2.2%	typical

multilingual transcription▂··▃▄█▆▅▄▂·▂ #7 of 36 * · 7.9% of its score · 2 tests

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

test	result	vs the others in its run	difficulty	counts	why
Open ASR multilingual WER	4.8% ✎	▂··▂▃▂█▃█▂▁▁ beats 20/29	avg 8.9%	▇▇▇ 7.5%	typical
Iberian ASR benchmark · 7-language macro WER (eu, ca, gl, pt, es + de, tr)	19.3% ✎	█▃▃··▅▃▃··▃▃ beats 7/10	avg 38%	▇ 0.4%	noisy

other tests54 not in a task · 90% of its score

test	result	vs the others in its run	difficulty	counts	why
LibriSpeech test-other WER +4	3.6% ✎	███·····██·█ beats 4/5	avg 5.7%	▇▇ 5.9%	several runs
LibriSpeech test-clean WER +5	1.8% ✎	▁···▁▁▂·▄▄██ beats 35/56	avg 2.3%	▇▇ 4.7%	blunt · several runs
AfriSpeech-MultiBench · 7-region African-accent mean WER	21.2% ✎	▂····▂···▂▆█ beats 11/11	avg 46%	▇▇ 4.2%	typical
SPGISpeech test WER +1	3.1% ✎	▄▄··▄██··▄██ beats 5/11	avg 5.3%	▇▇ 4.0%	sharp · several runs
GigaSpeech test WER +2	11.6% ✎	▁····▁··▁▄█▂ beats 12/17	avg 15%	▇ 3.7%	several runs
AfriSpeech-MultiBench · Parliament (legal domain) WER	20.0% ✎	▂▂····▂·▅·▃█ beats 8/12	avg 52%	▇ 3.6%	sharp
Bangla STT benchmark · BanSpeech 1,001-clip batch CER	30.7% ✎	█·▄··▄··▄·▄▄ beats 0/6	avg 37%	▇ 2.9%	sharp
FLEURS English test WER +5	4.0% ✎	▁··▁▁▁▁▁▁▃█▁ beats 51/56	avg 5.8%	▇ 2.6%	blunt · several runs

46 more tests

test	result	vs the others in its run	difficulty	counts	why
Mega-ASR adverse-condition suite · mean WER (CHiME-4, VOiCES, NOIZEUS)	10.7% ✎	▄····█▄▄···▄ beats 4/5	avg 19%	▇ 2.5%	typical
Open ASR English short-form WER (March 2026)	7.4% ✎	▁▁▁▃··▁▁▂██▆ beats 49/86	avg 8.2%	▇ 2.2%	blunt
VoxPopuli English test WER +2	7.2% ✎	▄··▄··▄···█▄ beats 1/5	avg 7.4%	▇ 2.2%	several runs
VoiceCodeBench · batch canonical token/entity match (CTEM)	87.6% ✎	▄···▄··▄▄▄·█ beats 2/6	easy · avg 77%	▇ 2.0%	typical
Common Voice English WER (version unspecified) +1	9.9% ✎	█··▄·······▄ beats 1/3	avg 13%	▇ 1.7%	several runs
AMI WER (Open ASR / ESB test split)	16.0% ✎	▃█▃▃·····▅·▃ beats 3/8	avg 24%	▇ 1.7%	typical
SwitchBoard WER	13.1% ✎	▄······▄···█ beats 1/3	avg 22%	▇ 1.7%	typical
Multilingual LibriSpeech French test WER	7.2% ✎	█·······██·█ beats 2/3	avg 11%	▇ 1.7%	typical
AfriCodeSwitch English-Yoruba code-switched speech · WER (cap2000)	71.2% ✎	█··█·█···█·█ beats 1/4	avg 131%	▇ 1.7%	typical
Earnings-22 WER (short-form, Open ASR / ESB test split)	11.3% ✎	▃······██▃·▃ beats 5/8	avg 18%	▇ 1.6%	typical
Omnilingual ASR paper · FLEURS-81 test mean CER (Whisper-covered languages)	22.6% ✎	▅···▃···▃▃██ beats 2/10	avg 13%	▇ 1.6%	sharp · noisy
Dutch ASR Leaderboard · average WER (FLEURS, VoxPopuli, MLS nl)	13.0% ✎	▄··▄█···█··▄ beats 0/6	avg 16%	▇ 1.6%	typical
FLEURS Mandarin test error rate +1	8.0% ✎	▂▂··▂▄▂▂·▆██ beats 15/17	avg 12%	▇ 1.6%	several runs
WenetSpeech test_meeting error rate +1	19.1% ✎	█··█······██ beats 1/3	avg 21%	▇ 1.6%	sharp · noisy · several runs
CHiME-4 WER	10.9% ✎	█··█·█·····█ beats 1/3	avg 18%	▇ 1.6%	typical
LiteASR paper · ESB 8-set average WER	10.1% ✎	▂·······▂··█ beats 6/6	avg 13%	▇ 1.5%	typical
Earnings-21 10-minute long-form WER	9.9% ✎	█···█·····██ beats 1/3	avg 16%	▇ 1.5%	typical
Earnings-22 10-minute long-form WER	13.1% ✎	█····█·█···█ beats 0/3	avg 21%	▇ 1.4%	typical
Multilingual LibriSpeech English test WER	4.9% ✎	█······█·█·█ beats 2/3	avg 7.7%	▇ 1.4%	typical
CoVoST 2 source speech, 13 languages, mean ASR WER (NVIDIA set)	10.3% ✎	█·····█····█ beats 1/2	avg 17%	▇ 1.3%	small run
FLEURS 25 European languages, mean WER (NVIDIA set)	12.6% ✎	██·········█ beats 0/2	avg 18%	▇ 1.3%	small run
WenetSpeech test_net error rate	9.9% ✎	█···█·····██ beats 1/3	avg 13%	▇ 1.3%	typical
Russian ASR Leaderboard (N. Shmyrev test sets) · 11-dataset mean WER	15.4% ✎	█·····█····█ beats 1/2	avg 25%	▇ 1.3%	small run
AISHELL-2 test error rate	5.1% ✎	█··█·····█·█ beats 0/3	avg 5.7%	▇ 1.3%	typical
GigaSpeechBench · 7 Arabic dialect WER mean	57.9% ✎	█··█·██··█·█ beats 0/5	avg 79%	▇ 1.3%	typical
Multilingual LibriSpeech 6 languages, mean WER (NVIDIA set)	8.0% ✎	█·▄········▄ beats 0/3	avg 11%	▇ 1.3%	typical
VoiceCodeBench · batch WER (300 English workplace segments)	5.8% ✎	▄·▄·▄·▄█···▄ beats 4/6	avg 12%	▇ 1.2%	typical
Multilingual LibriSpeech German test WER	13.2% ✎	█····█···█·█ beats 0/3	avg 13%	▇ 1.2%	typical
Omnilingual ASR paper · Common Voice 22 (76 languages) test mean CER	55.6% ✎	▂▂▃····▂▆··█ beats 2/10	avg 21%	▇ 1.1%	sharp · noisy
GigaSpeechBench · 12 English domains entity B-WER mean	14.7% ✎	▄·······▄·█▄ beats 3/4	avg 28%	▇ 1.1%	typical
PESTE · Persian FLEURS (fa_ir) corpus CER	6.0% ✎	▄··▄·▄▄▄··██ beats 8/8	avg 12%	▇ 1.1%	typical
Open ASR Leaderboard – Chilean Spanish · 3-dataset average WER	8.3% ✎	▂▂·····▂▂·█▂ beats 6/8	avg 13%	▇ 1.1%	typical
Distil-Whisper · 4-set out-of-distribution short-form mean WER	8.4% ✎	█·█··█·█···█ beats 4/4	avg 10%	▇ 1.0%	typical
BOTNOI report · FLEURS Thai CER	8.8% ✎	█··█·······█ beats 1/2	avg 17%	▇ 1.0%	small run
GigaSpeechBench · 6 English accents WER mean	16.3% ✎	▄······▄··▄█ beats 3/4	avg 36%	▇ 0.9%	typical
Common Voice Mandarin error rate (version unspecified)	12.9% ✎	█·······██·█ beats 0/3	avg 11%	▇ 0.9%	typical
Danish ASR Leaderboard · mean WER over 5 Danish test sets (RyeAI)	23.9% ✎	▂▂···▃▂▅█▃▂▂ beats 12/16	avg 32%	▇ 0.9%	blunt
BOTNOI Thai telephony · CER	48.4% ✎	█·······█··█ beats 0/2	avg 52%	▇ 0.9%	small run
Betthupferl (German dialect ASR benchmark) · Standard German audio → Standard German transcript WER	8.8% ✎	▄·█·▄▄▄▄·█▄▄ beats 10/10	avg 15%	▇ 0.7%	blunt
Open Universal Arabic ASR Leaderboard · 6-test-set average WER	36.9% ✎	▃·▂▂▅█▅▃▂▅▂▂ beats 13/22	avg 40%	▇ 0.6%	blunt
Picovoice STT benchmark · English batch 4-set mean WER	5.7% ✎	▄···▄···▄··█ beats 4/4	avg 5.7%	▇ 0.6%	blunt
TurkASR-Bench (7 Turkish sources, 5,200 utterances) · corpus WER	12.8% ✎	▃··▃▃··▅··█▅ beats 9/9	avg 17%	▇ 0.5%	blunt
Quranic ASR Leaderboard · held-out recitation (600 clips) overall WER	10.1% ✎	▃·▅·▅▃▃·▃▃▃█ beats 12/12	avg 24%	▇ 0.4%	blunt
Betthupferl (German dialect ASR benchmark) · dialect audio → Standard German transcript WER	30.7% ✎	▃█····▃▅▃·▅▃ beats 10/10	avg 41%	▇ 0.3%	blunt
Speech Accessibility Project dysarthric speech · zero-shot WER (rated subset, N=11,218)	14.4% ✎	▂·▂··▂····▆█ beats 7/9	avg 18%	▇ 0.3%	blunt
TED-LIUM test WER (short-form) +1	6.8% ✎	▁·▁▇█▃·▁▁▃█▄ beats 20/27	avg 5.3%	▇ 0.1%	blunt · several runs

$ where to run it

Through an API

api · Groq · $0.0018/audio min · console.groq.com ↗ · price sourceUSD per input audio minute, converted from the rounded LiteLLM per-second rate. 10-second minimum charge per request; published tariff $0.111/hour. Product: Batch/file speech transcription. API model: whisper-large-v3. Settings: streaming=no, minimum billed seconds=10. Price basis: rounded catalogue rate (not a new estimate of speech length). Currency: USD.

Other ways to run it

version

Run it as published · needs ~4.22 GB · from github.com/openai (official)
uvx --from openai-whisper whisper audio.mp3 --model large-v3

Official weights and documentation: huggingface.co/openai ↗

$ how the score was worked out · Mistral AI — Voxtral technica… 12%, Open ASR Leaderboard 12%, Qwen/Qwen3-ASR-1.7B model card 11% and 28 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

source	results	share of the score	licence
Mistral AI — Voxtral technical report	10	▇▇▇▇ 12%	CC BY 4.0
Open ASR Leaderboard	3	▇▇▇▇ 12%	Apache-2.0
Qwen/Qwen3-ASR-1.7B model card	13	▇▇▇ 11%	Apache-2.0
distil-whisper/distil-large-v3.5 model card	7	▇▇▇ 9%	MIT
AfriSpeech-MultiBench (arXiv:2511.14255v1) · African-accented English ASR benchmark suite	2	▇▇ 8%	CC BY 4.0
transcribe.cpp model catalog · LibriSpeech test-clean WER of validated full-precision ports	3	▇▇ 5%	MIT
LiteASR (arXiv:2502.20583v2) · ESB WER of Whisper and LiteASR-compressed encoders	3	▇ 4%	CC BY 4.0
Canary-1B-v2 & Parakeet-TDT-0.6B-v3: Efficient and High-Performance Models for Multilingual ASR and AST (NVIDIA, arXiv:2509.14128v2)	3	▇ 4%	CC BY-SA 4.0
microsoft/VibeVoice-ASR-BitNet model card (Accuracy table)	4	▇ 3%	MIT
GigaSpeechBench: A Real-World Multilingual Speech-to-Text Benchmark	3	▇ 3%	CC BY 4.0
VoiceCodeBench (arXiv:2608.28916v2) · structured-token recovery ASR benchmark	2	▇ 3%	CC BY 4.0
Bangla Speech-to-Text Benchmark v1.0 (Riverborn)	1	▇ 3%	MIT
Omnilingual ASR: Open-Source Multilingual Speech Recognition for 1600+ Languages (arXiv:2511.09690v1)	2	▇ 3%	CC BY-SA 4.0
Mega-ASR (arXiv:2605.19833v1) · adverse-condition ASR baselines (CHiME-4, VOiCES, NOIZEUS)	1	▇ 2%	CC BY 4.0
BOTNOI Thai telephony ASR report (arXiv:2608.24916v1) · baseline CERs	2	▇ 2%	CC BY 4.0
zai-org/GLM-ASR README benchmark table (GLM-ASR-Nano vs open/closed ASR)	4	▇ 2%	Apache-2.0
Switch-aware evaluation of ASR on English-Yoruba code-switched speech (arXiv:2609.11786v1)	1	▇ 2%	CC BY 4.0
Dutch ASR Leaderboard (tvosch) · committed result files	1	▇ 2%	Apache-2.0
Russian ASR Leaderboard (ESpeech; benchmark by Nikolay Shmyrev)	1	▇ 1%	MIT
PESTE · Persian Speech to Text benchmark (fleurs-fa-ir-v1)	1	▇ 1%	Apache-2.0
ai-sage/GigaAM-Multilingual model card (Model Performance table)	2	▇ 1%	MIT
Open ASR Leaderboard – Chilean Spanish (idsudd / Alonso Astroza)	1	▇ 1%	Apache-2.0
huggingface/distil-whisper README (Distil-Whisper headline WER table)	1	▇ 1%	MIT
German dialect ASR Benchmark (Betthupferl; kamaal-08/bavarian-asr-leaderboard)	2	▇ 1%	MIT
Danish Speech-to-Text Leaderboard (RyeAI) · results dataset	1	▇ 1%	MIT
Open Universal Arabic ASR Leaderboard (Elm Research Center)	1	▇ 1%	Apache-2.0
Picovoice speech-to-text-benchmark · English batch-engine WER (README)	1	▇ 1%	Apache-2.0
Turkish ASR Leaderboard (TurkASR-Bench, yagmurtuncer)	1	▇ 1%	MIT
Quranic ASR Leaderboard (Quran-Lab, held-out recitation set)	1	▇ 0%	Apache-2.0
Benchmarking Automatic Speech Recognition Tools for Iberian Languages (arXiv:2609.36920v1), Table 3	1	▇ 0%	CC BY 4.0
Dysarthric ASR with clinical context on the Speech Accessibility Project (arXiv:2605.02782v2)	1	▇ 0%	CC BY 4.0
prior · pull toward the middle		▇ 0%
weighted average		= 56.7

$ quantization evidence · 4 measurements

How much of the full-precision result a quantized version kept, measured within one study. Rows from the same study share a table and aren't independent.

precision	test	full → quantized	kept
FP8 dynamic (llm-compressor FP8_DYNAMIC; table column header says "w8a8")	LibriSpeech test-clean WER (lmms-eval `librispeech`, vLLM whisper backend, batch 64)	0.0217 → 0.021	103.6%	source ↗
FP8 dynamic (llm-compressor FP8_DYNAMIC; table column header says "w8a8")	LibriSpeech test-other WER (lmms-eval `librispeech`, vLLM whisper backend, batch 64)	0.039 → 0.0396	98.5%	source ↗
INT8 W8A8 (SmoothQuant + GPTQ via llm-compressor)	LibriSpeech test-clean WER (lmms-eval `librispeech`, vLLM whisper backend, batch 64)	0.0217 → 0.0207	104.7%	source ↗
INT8 W8A8 (SmoothQuant + GPTQ via llm-compressor)	LibriSpeech test-other WER (lmms-eval `librispeech`, vLLM whisper backend, batch 64)	0.039 → 0.039	100.2%	source ↗

$ improve this page

Update model information Add a place to run it Add a test result Propose a benchmark Correct a benchmark score Edit 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 2026-10-11 02:20 UTC · sources and licences · json