Models
Every model runs on your machine. They are downloaded, chosen and deleted in the app's settings: you stay in control of disk space. Speeds are measured, not promised.
One voice for every language. Kokoro installs itself on first launch; the system voice reads in the meantime.
| Model | Size | Languages | Speed |
|---|---|---|---|
| Kokoro 82M · MLX | 342 MB | French, English | 15× faster than speech on an M5 — busy ~6% of the time |
| Kokoro 82M · int8 | 108 MB | French, English | 8 to 10× faster than speech |
| Qwen3-TTS 0.6B · French voice | 2.0 GB | 10 languages, including French and English | Busy 76% of the time at 1× on an M5 — choppy above 1× |
| MagpieTTS 357M | 302 MB | 9 languages, including French and English | Busy 69% of the time at 1× on an M5 — choppy above 1× |
| System voice | 0 MB | The system's languages | Instant |
The dysflow voice. Natural, light, computed by the Mac's GPU. Installed automatically on first launch.
The same voice, compressed for the CPU. It is the one that reads in the terminal, on Linux and on Mac.
A fixed French voice, with a timbre drawn from the SIWIS corpus. More expressive, heavier.
NVIDIA's multilingual model, with five English voices. French is read as is.
The macOS voice, or espeak-ng on Linux. Nothing to download: it reads while Kokoro installs.
In the macOS app: a shortcut opens dictation, and the text is typed into the app in front. No model is imposed; a microphone test bench measures accuracy before you choose.
| Model | Size | Languages | Speed |
|---|---|---|---|
| Parakeet v3 | 2.5 GB | 25 European languages | Re-read every 1.5 s |
| Phonon-2 | 164 MB | English (French less reliable) | 8 s of audio transcribed in 0.12 s on an M5 |
| Whisper tiny | 23 MB | 99 languages | Re-read every 1.5 s |
| Whisper small | 140 MB | 99 languages | Re-read every 1.5 s |
| Whisper large-v3-turbo | 464 MB | 99 languages | Re-read every 1.5 s |
| Canary 1b-v2 | 1.1 GB | 25 European languages | Re-read every 1.5 s |
| Voxtral Realtime | 3.1 GB | 13 languages | As you speak |
| macOS Dictation | 0 MB | macOS's languages | As you speak |
Recognizes twenty-five European languages quickly and well. The default choice for dictating in French.
Parakeet v3 retrained for English and shrunk to 164 MB, almost as accurate. French works, less reliably.
The smallest Whisper: fast, for a short text or a modest machine.
A good trade-off between size and accuracy, in ninety-nine languages.
The most accurate Whisper in the list, at the cost of a heavier download.
Recognition and translation for twenty-five European languages.
Writes as you speak, without waiting for the end of the sentence.
macOS 26's on-device recognition. Nothing to download, and always there as a fallback.
For images, screenshots and scanned PDFs, dysflow bundles Tesseract with French and English: nothing to install. Pages are recognized in parallel and the result is remembered, so the second opening is instant.
Measured on an M5 Mac. “Busy x% of the time at 1×”: the share of playback time spent computing the voice; above 100%, playback stutters.