Results
Whisper Large V3 Turbo accuracy and speed
Whisper Large V3 Turbo is OpenAI's distilled Whisper, running locally through whisper.cpp. For a long time it was the default local model in Superwhisper, and it is still a sensible pick on a machine that cannot spare the memory for Cohere Transcribe.
Apple M4
70
Blended score
10.9%
Word error rate
1.23 s
Wait for the text
Macro average over 8 datasets on Apple M4
Apple M4 Max
72
Blended score
11.1%
Word error rate
0.34 s
Wait for the text
Macro average over 3 datasets on Apple M4 Max
On Apple M4
| Dataset | Score ↑ | WER ↓ | Speed ↑ | Recall ↑ | F-score ↑ | Response ↓ |
|---|---|---|---|---|---|---|
| AMI Meetingssw-app-coverage-esb-ami-v2-20260802 | 62 | 14.3% | 10× | — | — | 1.23 s |
| Common Voice, Australian Englishsw-app-coverage-cv24-australia-v2-20260802 | 86 | 4.7% | 4.4× | — | — | 1.23 s |
| Common Voice, spontaneous Englishsw-app-coverage-cv-spontaneous4-v2-20260802 | 50 | 19.0% | 8.8× | — | — | 1.23 s |
| Earnings 22sw-app-coverage-esb-earnings22-v2-20260802 | 75 | 8.9% | 15× | — | — | 1.23 s |
| Earnings 22 (with vocabulary)sw-app-main-contextual-earnings22-20260802 | 51 | 18.6% | 13× | 61.7% | 76.1% | 1.23 s |
| LibriSpeech Othersw-app-main-esb-librispeech-other-20260802 | 89 | 3.3% | 14× | — | — | 1.23 s |
| Loquacioussw-app-coverage-loquacious-v2-20260802 | 78 | 7.9% | 7.8× | — | — | 1.23 s |
| Phonetic vocabularysw-app-main-phonetic-vocab-20260802 | 72 | 10.2% | 2.6× | 71.4% | 74.6% | 1.23 s |
Also measured on Apple M4 Max
| Dataset | Score ↑ | WER ↓ | Speed ↑ | Recall ↑ | F-score ↑ | Response ↓ |
|---|---|---|---|---|---|---|
| Earnings 22 (with vocabulary)20260731-175829 | 49 | 20.4% | 46× | 73.9% | 83.7% | 0.34 s |
| LibriSpeech Other20260721-211510 | 91 | 3.5% | 53× | — | — | 0.34 s |
| Phonetic vocabulary20260731-175829 | 77 | 9.3% | 9.3× | 68.6% | 73.8% | 0.34 s |
Updated August 27, 2026 from bench commit 5344cab. The second line of each row is the run that produced it.
Method
How this was measured
Local results come from driving the real Superwhisper app over each clip, so they reflect the engine and decode settings that ship in the product rather than a separate implementation.
Word error rate counts insertions, deletions and substitutions against a human transcript, so lower is better. Speed is a multiple of real time, so higher is better. The full method and every other model is on the benchmarks page.
Compare
Other models in the benchmark
Same audio, same machines, same method.
Support