Results
Cohere Transcribe accuracy and speed
Cohere Transcribe runs on your machine through MLX at 4-bit precision. It is the most accurate offline option we ship, and the one to pick when audio cannot leave the laptop.
Quantising to 4-bit costs a little accuracy against the full-precision weights and buys a lot of memory back. These numbers are for the quantised build we actually ship.
Apple M4
79
Blended score
8.4%
Word error rate
0.45 s
Wait for the text
Macro average over 8 datasets on Apple M4
Apple M4 Max
78
Blended score
8.7%
Word error rate
0.18 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 | 81 | 7.4% | 25× | — | — | 0.45 s |
| Common Voice, Australian Englishsw-app-coverage-cv24-australia-v2-20260802 | 93 | 2.8% | 21× | — | — | 0.45 s |
| Common Voice, spontaneous Englishsw-app-coverage-cv-spontaneous4-v2-20260802 | 54 | 18.5% | 23× | — | — | 0.45 s |
| Earnings 22sw-app-coverage-esb-earnings22-v2-20260802 | 80 | 8.0% | 26× | — | — | 0.45 s |
| Earnings 22 (with vocabulary)sw-app-main-contextual-earnings22-20260802 | 66 | 13.8% | 29× | 69.1% | 81.3% | 0.45 s |
| LibriSpeech Othersw-app-main-esb-librispeech-other-20260802 | 94 | 2.2% | 26× | — | — | 0.45 s |
| Loquacioussw-app-coverage-loquacious-v2-20260802 | 87 | 5.2% | 23× | — | — | 0.45 s |
| Phonetic vocabularysw-app-main-phonetic-vocab-20260802 | 78 | 8.9% | 16× | 68.6% | 72.7% | 0.45 s |
Also measured on Apple M4 Max
| Dataset | Score ↑ | WER ↓ | Speed ↑ | Recall ↑ | F-score ↑ | Response ↓ |
|---|---|---|---|---|---|---|
| Earnings 22 (with vocabulary)20260731-175829 | 63 | 15.0% | 79× | 82.4% | 89.1% | 0.18 s |
| LibriSpeech Other20260721-211510 | 94 | 2.2% | 80× | — | — | 0.18 s |
| Phonetic vocabulary20260731-175829 | 78 | 9.0% | 33× | 77.1% | 79.4% | 0.18 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