voiceloop: fastest browser voice agent loop goes open source
TODOforAI has open-sourced voiceloop, a zero-dependency JavaScript library that runs the full VAD → STT → LLM → TTS loop in the browser. In a 30-turn benchmark, the Deepgram Flux plus ElevenLabs flash setup hit an 862ms median voice-to-voice latency with zero self-interruptions under echo coupling.
- 862ms median voice-to-voice latency, 1067ms p95 with Deepgram + ElevenLabs flash
- Fully local Piper TTS path runs with no cloud or keys at 974ms
- 0 self-interruptions in 30 echo-coupled turns vs 17–20 for OpenAI Realtime and Pipecat
- Supports any OpenAI-compatible LLM, four STT providers and swappable TTS
Read next
Software