Voice models
Switch engines under Settings → Voice models. First launch defaults to Apple on-device speech in most cases.
setup-03-models.webp)What Free allows
Section titled “What Free allows”LicenseManager.isASRProviderAllowed:
| Free | Trial / Pro |
|---|---|
| Apple (on-device) | Full catalog |
| Doubao (BYOK cloud) | Bailian, MiMo, Azure, Gemini, OpenAI Speech, Groq, ElevenLabs, Soniox, NVIDIA Speech NIM, OpenAI-compatible / self-hosted gateways, … |
| On-device SenseVoice-Small (sherpa-onnx · zh/yue/en/ja/ko) | |
| On-device Whisper Large-v3 Turbo (sherpa-onnx · 100+ languages) | |
| On-device Fun-ASR-Nano (sherpa-onnx · zh/en/ja + dialects) | |
| On-device Parakeet TDT 0.6B v3 (sherpa-onnx · EN + 25 EU) |
Pointing Base URL at localhost does not unlock multi-ASR on Free — OpenAI-compatible / faster-whisper still need Trial or Pro. The four built-in on-device engines (SenseVoice / Whisper Turbo / Fun-ASR-Nano / Parakeet) are Free.
Common engines
Section titled “Common engines”| Kind | Examples | Network | Plan |
|---|---|---|---|
| System on-device | Apple Speech / SpeechAnalyzer | Local (system may download language packs) | Free ✓; no cloud key |
| Built-in on-device (sherpa-onnx) | SenseVoice, Whisper Turbo, Fun-ASR-Nano, Parakeet TDT | Offline after first weight download | Free ✓; no API key |
| Regional cloud | Doubao, Bailian, MiMo | Yes | Free: Doubao only; Bailian / MiMo → Trial/Pro |
| General cloud | OpenAI Speech, Azure (MAI-Transcribe 2), Gemini, Groq, ElevenLabs, Soniox | Yes | BYOK; Trial/Pro |
| Compatible / self-host | OpenAI-compatible, faster-whisper | Depends | localhost OK; still Trial/Pro |
| NVIDIA cloud | Speech NIM (cloud Parakeet, …) | Yes / self-host | Trial/Pro (distinct from on-device Parakeet) |
Full list is whatever the in-app catalog shows. Built-in on-device models are batch (infer after release, not word-by-word streaming). First use downloads roughly hundreds of MB up to ~1.5 GB of runtime/weights.
API keys
Section titled “API keys”Stored in the macOS Keychain. Audio goes only to providers you configure (on-device engines never upload audio). See BYOK.
File transcription
Section titled “File transcription”File import can use a separate model from live voice typing. The same Free/Pro provider gate applies (Free: Apple / Doubao / four built-in on-device engines). See File transcription.
Local Only Mode
Section titled “Local Only Mode”Public-cloud endpoints fail when Local Only is on. Apple and the four built-in on-device engines are always allowed; other engines need loopback Base URL / WS and the plan gate above.
Proper nouns
Section titled “Proper nouns”Use each engine’s context / prompt / dictionary fields, then word replacements as a fallback.