🎙️ Voice Setup — CrustAI

CrustAI's voice pipeline runs entirely offline. No audio ever leaves your device.

Architecture

Mic / Phone → Whisper.cpp (STT) → LLM → Piper TTS → Speaker / Phone
                    ↑
             openWakeWord
            "Hey CrustAI"

1 — Install Whisper.cpp

git clone https://github.com/ggerganov/whisper.cpp
cd whisper.cpp
make

# Download the small model (~240 MB, good for Portuguese)
bash models/download-ggml-model.sh small

Available model sizes (speed vs accuracy tradeoff):

Model	Size	Speed	Accuracy
tiny	75 MB	⚡⚡⚡⚡	★★☆☆
base	140 MB	⚡⚡⚡	★★★☆
small	240 MB	⚡⚡	★★★★
medium	1.4 GB	⚡	★★★★★

2 — Install Piper TTS

pip install piper-tts

# Download a Portuguese voice
piper --model pt_BR-faber-medium --download-dir ./voices --update-voices

Available Portuguese voices: - pt_BR-faber-medium — Brazilian Portuguese, natural male voice - pt_PT-tugão-medium — European Portuguese

Browse all voices at: https://huggingface.co/rhasspy/piper-voices

3 — Enable in config

# config/config.yml
voice:
  enabled: true
  port: 8765
  wake_word: "hey crustai"
  stt_model: small
  stt_language: pt
  tts_voice: pt_BR-faber-medium

4 — Mobile access (iOS & Android)

CrustAI exposes a WebSocket on port 8765. Connect your phone using:

With Tailscale (recommended):

wss://crustai.your-tailnet.ts.net:8765/voice

iOS Shortcut: 1. Install Tailscale on both devices 2. Create a Shortcut: Record Audio → POST to CrustAI WebSocket → Play response

Android (Tasker): 1. Install Tasker + HTTP Request plugin 2. Create a task: Record → WebSocket send → Play audio response

Wake Word

CrustAI uses openWakeWord to listen for "Hey CrustAI" without consuming resources.

pip install openwakeword

You can train a custom wake phrase — see the openWakeWord docs.