Lyte Voice is a self-hosted alternative to managed voice-AI platforms — own the orchestration layer, swap providers per agent, and pay only for what you actually use. Built first for internal agency work, architected for multi-tenant SaaS from day one.
Lyte Voice speaking — how can I route your call today?
I need to reschedule tomorrow's 2 PM appointment.
Sure — I have openings Thursday at 10 AM or 3 PM. Which works?
Thursday at 3 works.
Signal chain
Audio in
WebSocket frames — browser or Twilio Media Stream
VAD + STT
Detect turn end, transcribe in near-real-time
LLM + RAG
Reason with context, retrieve from knowledge base
TTS
Stream synthesized speech back, frame by frame
No SFU. No SIP stack. No per-minute tax.
One pipeline, one protocol — raw audio over WebSocket — speaks to both your browser widget and the phone network.
Tune persona, provider, voice, and turn-taking behavior with live preview — no YAML, no redeploys.
Drop in PDFs and docs. They're chunked, embedded, and retrieved on every turn via pgvector.
Answer inbound lines or trigger outbound campaigns — the same agent, the same pipeline, your number.
Every call lands with duration, transcript, and status — rolled up into usage you can actually act on.