whatmodel.sh / stt / Whisper large-v3-turbo
Whisper large-v3-turbo
↗ share
+ contribute
Whisper large-v3-turbo is an open-weights speech-to-text model from OpenAI, released October 2024. Its published size is 0.81B parameters. Licence: mit.
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-turbo, out of 176 with a score, then more from OpenAI.
| # | model | score | where | ||
|---|---|---|---|---|---|
| 59 |  Omnilingual ASR LLM 7B v2 | 58±3 | local | ||
| 60 |  Distil-Whisper distil-large-v3.5 | 58±1 | local | ||
| 61 | Whisper large-v3-turbo | 57±1 | api $0.0005/audio min +local | ||
| 62 | Omnilingual ASR LLM 3B | 57±2 | local | ||
| 63 |  Fun-ASR-MLT-Nano-2512 | 57±4 | 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 | ||
| 67 | Whisper large-v3 | 57±1 | api $0.0018/audio min +local | ||
see all speech-to-text models · find the ones that fit your machine
$ what it's good at
long-form transcription▁▁▄▁·▁▂█▄▁▁▁ #6 of 27 * · 4.9% of its score · 1 test
Beats Chirp 3, Parakeet TDT 0.6B v2, Parakeet TDT 1.1B here, though they rank higher overall. Trails Whisper large-v3 here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Open ASR English long-form WER | 8.7% ✎ | ▁▁▅▁·▃▅█▆▁▂▂ beats 22/30 | 4.9% |
multilingual transcription▂··▃▄█▆▅▄▂·▂ #15 of 36 * · 17% of its score · 2 tests
Beats Cohere Transcribe Arabic (07-2026), Qwen3-ASR-1.7B, Voxtral Mini Transcribe V2 here, though they rank higher overall. Trails Omnilingual ASR CTC 1B v2, SeamlessM4T v2 Large, Omnilingual ASR LLM 1B v2 here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Open ASR multilingual WER | 5.6% ✎ | ▂··▂▃▂█▃█▂▁▁ beats 16/29 | 17% |
| CoSE-E enterprise code-switching · 4 European pairs mean WER | 43.6% ✎ | █··██···█··█ beats 0/4 | — |
other tests28 not in a task · 78% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| LibriSpeech test-clean WER +1 | 2.0% ✎ | ▁···▁▁▂·▄▄██ beats 26/56 | 7.1% |
| Open ASR English short-form WER (March 2026) | 7.8% ✎ | ▁▁▁▃··▁▁▂██▆ beats 41/86 | 4.9% |
| SPGISpeech test WER | 3.0% ✎ | ▄▄··▄██··▄██ beats 4/11 | 4.7% |
| Meanwhile long-form WER | 5.0% ✎ | █·······██·█ beats 3/3 | 4.3% |
| TED-LIUM long-form WER | 3.1% ✎ | █······█·█·█ beats 3/3 | 4.0% |
| AMI WER (Open ASR / ESB test split) | 16.1% ✎ | ▃█▃▃·····▅·▃ beats 1/8 | 3.8% |
| Earnings-22 WER (short-form, Open ASR / ESB test split) | 11.6% ✎ | ▃······██▃·▃ beats 3/8 | 3.6% |
| LiteASR paper · ESB 8-set average WER | 10.1% ✎ | ▂·······▂··█ beats 4/6 | 3.3% |
20 more tests
$ where to run it
Through an API
api · Cloudflare · $0.0005/audio min · developers.cloudflare.com ↗ · price sourceCloudflare Workers AI published USD per input audio minute; file transcription, not WebSocket streaming. Plan base fees excluded. USD table is rounded; backend bills Neurons. Product: Workers AI file transcription. API model: @cf/openai/whisper-large-v3-turbo. Settings: streaming=no, plan=paid, free allocation=excluded. Price basis: published rounded USD/audio-minute. Currency: USD.
Other ways to run it
version
- Run it as published · needs ~2.44 GB · from github.com/openai (official)
uvx --from openai-whisper whisper audio.mp3 --model turbo
Official weights and documentation: huggingface.co/openai ↗
$ how the score was worked out · distil-whisper/distil-large-v… 36%, Open ASR Leaderboard 26%, transcribe.cpp model catalog… 11% and 10 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
not counted (1)
| CoSE-E: A Benchmark for Code-switched Speech Evaluation in Enterprise Settings (arXiv:2609.35645v1), Table 2 | CoSE-E enterprise code-switching · 4 European pairs mean WER [correct] | 43.6% | failed run: Source flags a decoding-mode failure, not an accuracy measurement: 'Run without a specified language ... it does not transcribe the mixed utterance at all: it detects a single language and translates the rest of the speech into English ... This is a decoding-mode failure ... so we report Whisper separately' (§5, Appendix D.3). |
$ published speed measurements · 6
Reference results for the hardware, engine and settings below. They aren't verified speeds of the downloads above or predictions for your machine.
| AMD Ryzen 7 9800X3D | onnx-default (precision not stated in table) | 3.9 x-realtime | ONNX Runtime CPUExecutionProvider · English suite; VoxPopuli en test; 100 samples used for warmup; per-waveform recognition timing; RTFx = summed audio duration / summed recognition time Exact published English table cell. Default precision is not silently relabeled fp32. TensorRT fp16 enables that provider setting; it does not assert every operation uses fp16. | source ↗ |
| AMD Ryzen 7 9800X3D | onnx-int8 | 5.4 x-realtime | ONNX Runtime CPUExecutionProvider · English suite; VoxPopuli en test; 100 samples used for warmup; per-waveform recognition timing; RTFx = summed audio duration / summed recognition time Exact published English table cell. Default precision is not silently relabeled fp32. TensorRT fp16 enables that provider setting; it does not assert every operation uses fp16. | source ↗ |
| NVIDIA GeForce RTX 5070 Ti | onnx-fp32 | 54.8 x-realtime | ONNX Runtime CUDAExecutionProvider · English suite; VoxPopuli en test; 100 samples used for warmup; per-waveform recognition timing; RTFx = summed audio duration / summed recognition time Exact published English table cell. Default precision is not silently relabeled fp32. TensorRT fp16 enables that provider setting; it does not assert every operation uses fp16. | source ↗ |
| NVIDIA GeForce RTX 5070 Ti | onnx-fp16 | 76 x-realtime | ONNX Runtime CUDAExecutionProvider · English suite; VoxPopuli en test; 100 samples used for warmup; per-waveform recognition timing; RTFx = summed audio duration / summed recognition time Exact published English table cell. Default precision is not silently relabeled fp32. TensorRT fp16 enables that provider setting; it does not assert every operation uses fp16. | source ↗ |
| NVIDIA T4 | onnx-fp32 | 9.2 x-realtime | ONNX Runtime CUDAExecutionProvider · English suite; VoxPopuli en test; 100 samples used for warmup; per-waveform recognition timing; RTFx = summed audio duration / summed recognition time Exact published English table cell. Default precision is not silently relabeled fp32. TensorRT fp16 enables that provider setting; it does not assert every operation uses fp16. | source ↗ |
| NVIDIA T4 | onnx-fp16 | 25.9 x-realtime | ONNX Runtime CUDAExecutionProvider · English suite; VoxPopuli en test; 100 samples used for warmup; per-waveform recognition timing; RTFx = summed audio duration / summed recognition time Exact published English table cell. Default precision is not silently relabeled fp32. TensorRT fp16 enables that provider setting; it does not assert every operation uses fp16. | source ↗ |
$ quantization evidence · 6 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) | 0.0227 → 0.0226 | 100.3% | source ↗ |
| FP8 dynamic (llm-compressor FP8_DYNAMIC; table column header says "w8a8") | LibriSpeech test-other WER (lmms-eval `librispeech`, vLLM whisper backend, batch 64) | 0.0424 → 0.0432 | 98.0% | source ↗ |
| INT4 W4A16 (GPTQ, weights only, group size 64 per the quantize.py command shown, actorder=weight) | LibriSpeech test-clean WER (lmms-eval `librispeech`, vLLM whisper backend, batch 64) | 0.0219 → 0.022 | 99.7% | source ↗ |
| INT4 W4A16 (GPTQ, weights only, group size 64 per the quantize.py command shown, actorder=weight) | LibriSpeech test-other WER (lmms-eval `librispeech`, vLLM whisper backend, batch 64) | 0.039 → 0.0404 | 96.5% | source ↗ |
| INT8 W8A8 (SmoothQuant + GPTQ via llm-compressor) | LibriSpeech test-clean WER (lmms-eval `librispeech`, vLLM whisper backend, batch 64) | 0.0227 → 0.0226 | 100.2% | source ↗ |
| INT8 W8A8 (SmoothQuant + GPTQ via llm-compressor) | LibriSpeech test-other WER (lmms-eval `librispeech`, vLLM whisper backend, batch 64) | 0.0424 → 0.0431 | 98.3% | source ↗ |
$ 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