v0.7 · Local Speech Engines · Optional Downloads
Speak any of 98 languages.
Korean, Japanese, Chinese, Hindi, Arabic, Vietnamese, Thai — and 91 others. All on your Mac. All free.
Pick yours.
A sample of the 98 supported languages.
Plus 74 more — including everything Parakeet covers.
Local engines, one app.
Parakeet for the everyday path. WhisperKit for broad languages. Cohere for local batch file transcription.
Parakeet
The default engine. Fastest local STT on Apple Silicon, sharpest on English.
- ▸ 25 European languages including English, Spanish, French, German, Italian, Portuguese, Russian
- ▸ 155× realtime on Apple Silicon
- ▸ ~2.5% word error rate
- ▸ ~465 MB per selected Parakeet build, downloaded during setup
WhisperKit
One-time download adds 73 more languages. Slower than Parakeet for English — but covers far more of the world.
- ▸ 98 languages including all of Parakeet's
- ▸ Korean, Japanese, Chinese, Hindi, Arabic, Vietnamese, Thai, and 73 more
- ▸ 632 MB one-time download
- ▸ Switch in Settings or pick per-call from the CLI
Cohere Transcribe
Local batch transcription for files when you want another model to compare. Best for final plain text, not live dictation.
- ▸ Runs locally after model download, with memory checks before use
- ▸ GPU/ANE compute setting for supported Macs
- ▸ Plain-text final transcript; no live preview or word timestamps
- ▸ Useful for batch file workflows, not the default dictation path
How to switch engines
Pick the local engine that matches the task. Optional CLI override per call.
Open Settings
Settings → Speech Recognition. Pick a speech engine and model.
Download once
Optional engines download their model files the first time. After that, core speech recognition runs on your Mac.
Pick a language
For WhisperKit, pick a default — or leave it on Auto to detect.
Power users: override the engine or language per call from the CLI.
$ macparakeet-cli transcribe interview.m4a \
--engine whisper --language ko --format jsonMore model choices. Same promise.
Switching engines changes the model, not the principle. Your audio still never leaves your Mac for core speech recognition. Optional AI summaries are separate and only use transcript text when you configure a provider.
Questions & answers
Why isn't WhisperKit the default?
Speed. Parakeet is faster on Apple Silicon and sharper on English. WhisperKit is the right choice when you need its languages, but for English speakers, Parakeet gets you to text first.
Do I need every engine?
No. Pick the engine that fits the job. Parakeet is the everyday default, WhisperKit is for broader language coverage, and Cohere Transcribe is an optional local batch engine for plain-text file transcription.
Is my audio still on-device with optional engines?
Yes. WhisperKit, Cohere Transcribe, and Parakeet all run locally after their model files are downloaded. No audio leaves your Mac for core speech recognition.
How accurate is WhisperKit on my language?
It varies by language. WhisperKit tracks Whisper's published benchmarks: strong on European languages and Mandarin, solid on Japanese and Korean, weaker on low-resource languages. Try it on a 30-second clip before you commit.
Can I download smaller or larger Whisper models?
The default WhisperKit option is large-v3-turbo, a 632 MB model with a strong size and quality trade-off on Apple Silicon. MacParakeet keeps the UI focused on the supported model choices rather than exposing every research variant.
Does dictation work in non-English languages?
Yes, with engines that support the live dictation path. Parakeet and WhisperKit can be selected for dictation; Cohere Transcribe is batch-oriented, so it does not provide live dictation preview.
Will Parakeet add more languages?
Likely. Parakeet TDT is actively maintained by NVIDIA's research team, and its v3 release added 25 European languages. New ones get picked up as they ship. WhisperKit covers the gap until then.
Try it in your language.
Free and open-source. Download MacParakeet, switch the engine in Settings, speak.
macOS 14.2+ · Apple Silicon · GPL-3.0