Voice (STT & TTS)
Ghast does voice both ways — listening (Whisper STT) and talking (Kokoro or Qwen TTS) — and runs both locally on your machine. No audio leaves the device for transcription or synthesis.
Speech-to-Text (Whisper)
Whisper is the local transcription engine. It runs in-process (via @fugood/whisper.node) on either CPU or Metal (Apple Silicon), at 16 kHz with automatic resampling.
Where you'll use it
- Composer microphone — press the mic in the composer to dictate; transcription streams into the input.
- Inbound voice messages — Telegram / Discord / WeChat voice clips are transcribed before the agent sees them.
- Voice skill — explicit recording / transcription invoked from chat.
Settings
Settings → Whisper:
- Backend — CPU or Metal (Apple Silicon only).
- Model variant — choose between size/quality tradeoffs.
- Language — auto-detect or pin a specific language (Whisper supports many; common ones in the picker).
The first time you use voice input, Ghast requests microphone permission and (on demand) downloads the chosen model.
Text-to-Speech (TTS)
Two local engines, picked by you in Settings → TTS.
| Engine | Strengths | Availability |
|---|---|---|
| Kokoro | 82M model, runs anywhere, supports multiple languages | All Macs |
| Qwen TTS (MLX) | Higher quality, especially for Mandarin | Apple Silicon only |
Where you'll hear it
- Voice replies in remote channels — when the user's message was voice, the agent's reply can be synthesised back to voice.
- Voice skill — explicit playback in chat.
- Optional read-aloud — for hands-free workflows.
Voice selection
Each engine ships several voices. The picker in Settings → TTS shows what's available; selection is per-engine. Voice changes apply immediately to the next utterance.
Bandwidth and footprint
- Whisper model files are ~50–500 MB depending on variant; downloaded once.
- TTS model files are similar in scale.
- Inference happens on-device. No per-second cost; no network round-trip.
Privacy
- Audio captured from the microphone is transcribed in-process and discarded. There is no audio log.
- Inbound voice clips from messaging bridges are transcribed and stored only as text in conversation history (unless the original message URL is also retained by the bridge).
- TTS output is rendered in-memory and either played directly or sent to the remote channel; no audio file is persisted unless you ask for one.
Failure modes
| Symptom | Likely cause | Action |
|---|---|---|
| Mic input fails | Permission not granted | System Settings → Privacy & Security → Microphone |
| Transcription quality is poor | Wrong language pinned, low-quality model variant | Settings → Whisper → switch to a larger variant or auto-detect language |
| TTS voice sounds wrong on Intel | Qwen is selected but not available | Settings → TTS → switch to Kokoro |
| Voice in Telegram is missing | The Telegram bridge isn't configured for voice IO | Settings → Channels → Telegram, enable voice |