Voice interface layer
VoiceForge — glass-box voice triage
Speak or type an IT issue and watch ASR → governed LLM → TTS replay under hard latency budgets. Architecture and live SLOs on the left, the product on the right.
Voice pipeline · phase replay
awaiting turnASR→LLM triage→TTS
ASR → governed LLM → TTS — each stage runs under a hard latency budget with graceful degradation.
Run a voice turn to replay ASR → LLM → TTS latency…
stages 3total n/aTTFT n/adegradation none
Stack status (details)
| Component | Status | Notes |
|---|---|---|
| Browser ASR | ✅ Live | Web Speech API — default on Render free tier |
| Server Whisper | 🔧 Optional | pip install -e ".[asr]" on GPU host |
| LLM triage | ✅ Mock / Ollama / DomainForge | Pairs with DomainForge S0→S4 ladder |
| Edge TTS | ✅ Server | Browser speechSynthesis fallback |
| WebSocket | ✅ /ws/voice | Phase events + result payload |