Voice to text on Mac, built in or with Koi
How to use voice to text on Mac with Apple's built-in Dictation, step by step, and when Koi does more: offline, 100 languages, unlimited on every plan.
The list
| Model | Made by | License | Size | Languages | Finds the language |
|---|---|---|---|---|---|
| Parakeet v3 (built in) | NVIDIA | CC BY 4.0 | 740 MB | 25 | Yes |
| Whisper Base | OpenAI | MIT | 85 MB | 99 | Yes |
| Whisper Small | OpenAI | MIT | 270 MB | 99 | Yes |
| Whisper Large v3 Turbo | OpenAI | MIT | 886 MB | 100 | Yes |
| Qwen3-ASR 1.7B | Alibaba Qwen | Apache 2.0 | 2.2 GB | 29 | Yes |
| Cohere Transcribe | Cohere | Apache 2.0 | 2.4 GB | 14 | No, you pick it |
| Granite Speech 4.1 | IBM | Apache 2.0 | 2.6 GB | 6 | No, you pick it |
All seven run on your Mac, so dictation, recorded calls and transcribed files never go online. Parakeet comes inside the app. The other six are free downloads, one click each in Settings, under Dictation, and each is checked before Koi uses it. Between them, Koi writes 100 languages.
The default
NVIDIA's Parakeet TDT 0.6B v3 is what Koi uses until you pick another. It's the fastest of the seven and works out by itself which of its 25 European languages you're speaking: Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish and Ukrainian. These are the languages Koi offers when you first set it up.
The downloads
Accuracy
| Model | Words wrong, LibriSpeech test-clean |
|---|---|
| Cohere Transcribe | 1.27% |
| Granite Speech 4.1 | 1.32% |
| Qwen3-ASR 1.7B | 1.62% |
| Parakeet v3 | 1.94% |
| Whisper Large v3 Turbo | 2.01% |
| Whisper Small | 3.33% |
| Whisper Base | 5.12% |
These are the scores published with each model's transcribe.cpp conversion, the same files Koi runs, on clean read English. Your accent, microphone and room matter more than the gaps between the top four, so try two and keep the one that gets your words right.
Choosing
Switch in Settings, under Dictation. The model you're using keeps working until the new one has downloaded. The bigger models run best on Apple silicon. The same model writes down your dictation (Fn + ⌥), your recorded calls and the files you drop in.
Languages
Afrikaans, Albanian, Amharic, Arabic, Armenian, Assamese, Azerbaijani, Bashkir, Basque, Belarusian, Bengali, Bosnian, Breton, Bulgarian, Burmese, Cantonese, Catalan, Chinese, Croatian, Czech, Danish, Dutch, English, Estonian, Faroese, Finnish, French, Galician, Georgian, German, Greek, Gujarati, Haitian Creole, Hausa, Hawaiian, Hebrew, Hindi, Hungarian, Icelandic, Indonesian, Italian, Japanese, Javanese, Kannada, Kazakh, Khmer, Korean, Lao, Latin, Latvian, Lingala, Lithuanian, Luxembourgish, Macedonian, Malagasy, Malay, Malayalam, Maltese, Maori, Marathi, Mongolian, Nepali, Norwegian, Norwegian Nynorsk, Occitan, Pashto, Persian, Polish, Portuguese, Punjabi, Romanian, Russian, Sanskrit, Serbian, Shona, Sindhi, Sinhala, Slovak, Slovenian, Somali, Spanish, Sundanese, Swahili, Swedish, Tagalog, Tajik, Tamil, Tatar, Telugu, Thai, Tibetan, Turkish, Turkmen, Ukrainian, Urdu, Uzbek, Vietnamese, Welsh, Yiddish, Yoruba.
Each comes from at least one of the models above. The language list in Settings shows the ones the model you're using knows. Some languages Whisper knows have far less training behind them, so expect more mistakes in them than in English or Spanish.
How to use voice to text on Mac with Apple's built-in Dictation, step by step, and when Koi does more: offline, 100 languages, unlimited on every plan.
The best Whisper apps for Mac in 2026, for dictation and for transcribing files: prices, free options, Intel support and which Whisper models each one runs.
The best apps for offline dictation on Mac: Koi, Apple Dictation, Superwhisper, MacWhisper, VoiceInk, Spokenly and more. What runs on device, and how to check.
Transcribe voice memos, mp3, m4a, mp4 and mov files on your Mac, offline, with timestamps you can click, a summary if you want one, and SRT or VTT export.
Support
Try Pro free for 3 days, card needed. Dictation and recorded calls are unlimited on every plan.