More Than One Way to Hear You
No single speech model is best at everything. Whisper knows many languages, Parakeet is very fast, Apple’s own engine needs no download, and some small models specialise in one family of languages. VocaMac includes all four kinds and runs every one of them on your Mac. You choose in Settings → Speech Model, and you can switch at any time.
The Engines
| Engine | Runs on | Good at | Models |
|---|---|---|---|
| Whisper (WhisperKit) | CoreML on Apple Silicon | The widest language coverage, translation to English | Tiny to Large v3, Distil variants, Voca Hinglish |
| Parakeet (NVIDIA, via FluidAudio) | CoreML on the Neural Engine | Speed; English and 25 European languages | Parakeet 110M, v2, v3, Ultra, Unified |
| Apple Speech | macOS 26 system engine | No model download; assets managed by macOS | Apple Speech |
| Specialized ONNX (sherpa-onnx) | CPU only | Compact models for specific languages and low-memory Macs | Moonshine, SenseVoice, GigaAM v3, Canary 180M Flash, Qwen3 ASR 0.6B |
Whisper
OpenAI’s Whisper, compiled to CoreML by WhisperKit, is the generalist. It has the most selectable languages, can translate speech into English, and comes in sizes from Tiny (77 MB) to Large v3 (1.6 GB). Voca Hinglish is a Whisper model fine-tuned to write Hindi speech in Roman script.
Parakeet
NVIDIA’s Parakeet models run on the Apple Neural Engine and are the fastest option in VocaMac. Parakeet 110M (227 MB), v2 and Unified are English-only; Unified adds punctuation, capitals and numbers as digits. Parakeet v3 and Ultra cover 25 European languages. Parakeet needs Apple Silicon, which every Mac VocaMac supports already has.
Apple Speech
On macOS 26, VocaMac can use Apple’s SpeechAnalyzer engine. There is nothing to download from VocaMac: macOS manages the language assets, and which languages are available is decided by macOS.
Specialized ONNX Models
These run on the CPU through sherpa-onnx and suit particular languages or Macs with little memory:
- Moonshine v2 Tiny and Base: English, for very low-RAM Macs (44 MB and 142 MB)
- SenseVoice: Chinese, Japanese, Korean, Cantonese and English
- GigaAM v3: Russian, with punctuation
- Canary 180M Flash: English, Spanish, German and French
- Qwen3 ASR 0.6B: 30 languages, including Hindi, Arabic and Thai
Which One Should I Use?
During setup, VocaMac asks which languages you speak and recommends a model:
| You mostly speak | Suggested starting point |
|---|---|
| English | Parakeet 110M |
| Russian | GigaAM v3 |
| Chinese, Japanese or Korean | SenseVoice |
| Spanish, German or French | Canary 180M Flash |
| Several languages | Whisper Small |
Tiny ships with the app, so dictation works before any download. All of these are starting points; Smart Model Selection explains how memory and accuracy trade off, and Custom Vocabulary works with every engine.
On-Device by Default
Every engine above transcribes on your Mac. A model needs an internet connection only for its first download (Apple Speech assets come from macOS). VocaMac also offers Custom Endpoint, an opt-in choice that sends each recording to a Whisper-compatible server you configure instead. It is never the default, and if you use it, keep the server on this Mac, on a trusted local network, behind Tailscale, or on HTTPS.
Ready to try Speech Engines?
Install VocaMac