Models
The best speech model for dictation on your Mac
Every model Codictate can run, measured on the same recordings on an Apple M4 Max. Pick by the language you dictate and the memory you can spare.
Which model to pick
Most accurate across languages
Whisper Large V3 got 94.0% of words right across all five test languages, the best of any model here, and peaks at 3.5 GB of RAM. Whisper Large V3 Q5_0 scores 94.0% on 1.5 GB, a tie on this many recordings, so it is the easier one on your Mac.
Fastest
Parakeet TDT 0.6B v3 makes you wait 11 ms for every second of audio you dictate, at 92.0% word accuracy. It covers 25 languages rather than the whole Whisper list. It does not translate.
Lightest on memory
The fastest model is also the lightest: Parakeet TDT 0.6B v3 peaks at 80 MB of RAM.
Danish
Edda v0.2 Q5_0 is the most accurate Danish model here, at 92.5% of words right. Wispr Flow is still more accurate on Danish than any model here; the head to head has the numbers.
Every model, side by side
Word accuracy is the share of words transcribed correctly, so higher is better. Wait is milliseconds per second of audio, so lower is better. A Danish-only model is scored on Danish alone.
| Model | Word accuracy | Wait per second of audio | Peak RAM | Download | Languages |
|---|---|---|---|---|---|
| Whisper Large V3 Turbo Q5_0 | 93.5% | 66 ms | 778 MB | 574 MB | All Whisper languages |
| Whisper Large V3 Q5_0 | 94.0% | 106 ms | 1.5 GB | 1.1 GB | All Whisper languages |
| Whisper Large V3 | 94.0% | 164 ms | 3.5 GB | 3.0 GB | All Whisper languages |
| Whisper Large V3 Turbo | 93.6% | 93 ms | 1.9 GB | 1.6 GB | All Whisper languages |
| Whisper Medium Q5_0 | 90.8% | 65 ms | 878 MB | 514 MB | All Whisper languages |
| Whisper Small Q5_1 | 80.7% | - | 383 MB | 181 MB | All Whisper languages |
| Parakeet TDT 0.6B v3 | 92.0% | 11 ms | 80 MB | 500 MB | 25 |
| Hviske v5 Tiny Q5_0 | 88.7% (Danish) | 18 ms | 327 MB | 181 MB | Danish |
| Edda v0.2 Q5_0 | 92.5% (Danish) | 53 ms | 787 MB | 547 MB | Danish |
Per-language figures, Wispr Flow and every other model we have run are on the benchmarks page. How the recordings are chosen and scored is in how we measure.
Run any of them in Codictate
Codictate is free and open source for Mac and Windows. Pick a model in Settings, press the shortcut, and the text lands in whatever app your cursor is in. Nothing leaves your machine.
Get Codictate