← All work

voiceAgent

A fully local, real-time voice assistant you can talk over: Parakeet listens, an LLM thinks, Kokoro speaks, SQLite remembers. Sub-second speech-to-speech with two-stage barge-in.

September 2026ShippedSource on GitHub ↗
voiceAgent screenshot

voiceAgent is a local, real-time voice assistant for Windows built around one constraint: it has to be fast enough that talking to it feels like a conversation, and it has to stop when you talk over it.

Audio from the mic goes through Silero VAD and an echo gate into a state machine. Parakeet TDT 0.6B transcribes each utterance (a CUDA + int8 CPU hybrid), Smart Turn v3.2 decides whether a pause is a real end of turn, and a LangGraph "brain" streams a reply with tools. Kokoro splits the streaming text into sentences and starts speaking on the first one.

Barge-in happens in two stages: 96 ms of speech ducks the TTS to 25%, 320 ms of sustained speech cuts it and cancels the model stream. Every stage is benchmarked and the numbers are in the repo's results/REPORT.md.

Hard parts

  • Holding mid-sentence pauses without adding latency at real turn ends (Smart Turn endpointing)
  • Stopping the agent interrupting itself: echo-gated, two-stage barge-in
  • Keeping first audio under half a second by streaming TTS sentence by sentence

Gallery

Live session in the Textual TUI: transcription, barge-in and tool use
Live session in the Textual TUI: transcription, barge-in and tool use

More work