Whisper
OpenAI · United States · Speech & audio
Global Open sourceFree tierIndependent test
Showing version: large-v3 · 2023-11-06 · Version lineage ↓
01
Overview
OpenAI's open-source general-purpose speech recognition family for multilingual transcription, speech translation and language identification, released under the MIT license and a de-facto open ASR baseline; OpenAI also serves it as a hosted API (whisper-1).
Release date2023-11-06
02 / Provenance
Field checks and source notes
When sources disagree on a field, every reading is listed here with its own citation. We do not take a midpoint, invent a missing figure, or decide which side is right — follow the citation and judge for your case.
These field-level notes are recorded verbatim in Chinese in the source data. Open the Chinese version
03 / Plain layer
Plain-language benchmarks
Naturalness Not disclosed
Accuracy (WER) 2.7% · lower is better
Speed / latency Not disclosed
Languages supported Not disclosed
Price / free Not disclosed开源权重免费自部署
04
Practical facts
| Context length | Not disclosed |
|---|---|
| Max output | Not disclosed |
| Parameters | Not disclosed |
| Input | Not disclosed |
| Output price | Not disclosed |
| CNY | Not disclosed |
| Free tier | Open weights, free to self-host. |
| Output speed | Not disclosed |
| Release date | |
| License | Not disclosed |
Show technical data
Technical benchmark table
| Benchmark | Score | Max / unit | Source type | Source | Test date | Conditions |
|---|---|---|---|---|---|---|
| ASR word error rate asr-wer | 2.7% | percent / 100 | Independent test | LibriSpeech test-clean（干净朗读英语）；越低越好。注：另有第三方称同测试集约 2.01%，数值随评测实现有出入，见 gaps |
| Version | Release date | Status | Openness | Parameters | What changed | Detail | |
|---|---|---|---|---|---|---|---|
| large-v3-turbo | Active | Open source | Not disclosed | Speed-optimised cut of large-v3: decoder pruned to 4 layers, ~809M parameters, several times faster at near-equal accuracy. | Detail | ||
| large-v3Showing version | Active | Open source | Not disclosed | Released at OpenAI DevDay 2023: 128 Mel bins, added Cantonese token, trained on ~5M hours including pseudo-labelled audio. | Detail | ||
| large-v2 | Superseded | Open source | Not disclosed | Improved training and regularisation over the original large model. | Detail | ||
| Whisper 首发（tiny→large 五档） | Superseded | Open source | Not disclosed | Open-sourced September 2022, trained on ~680k hours of weakly supervised audio, code and weights under MIT. | Detail |
Last updated: · Sources on detail page
06 / Compare