VoiceMesh

VoiceMesh is an open-source, production-inspired reliability lab for real-time voice AI infrastructure. It's a complete, self-hostable reference system showing how to build scalable, observable speech-to-speech pipelines — VAD → STT → LLM → TTS → transport — with the operational plumbing real deployments need: backpressure, event streaming, durable workflows, billing, and full observability.

Features

• End-to-end real-time pipeline: WebRTC VAD → STT → LLM → TTS → transport
• Backpressure handling via weighted in-memory queues
• Provider abstraction to swap STT / LLM / TTS backends
• Kafka event streaming with replay/durability; Temporal durable workflows
• PostgreSQL state + transactional outbox; optional ClickHouse analytics
• Usage-based billing metering (STT / LLM / TTS)
• Full observability: OpenTelemetry, Jaeger, Prometheus, Grafana
• Browser dashboard with mic input and PCM playback

Who Is It For?

• Engineers building or operating real-time voice AI systems
• Teams studying reliable, observable voice-agent infrastructure

Use Cases

• A reference architecture for production voice pipelines
• Experimenting with backpressure, replay, and durability
• Learning observability and billing for voice AI

---

Study production-grade voice AI infrastructure — VoiceMesh on GitHub.