Open-source speech-to-speech cuts voice agent history to exactly what you heard on interrupt

andimarafioti · x · 2026-11-18

The open-source speech-to-speech project (13.7k stars) by Hugging Face fixes a subtle voice agent issue: agents generate text faster than they speak, so interrupting one makes the model assume you heard the full reply. The new update trims the model's chat history to exactly what the user actually heard.

The project runs entirely locally with a default pipeline of Silero VAD → Parakeet TDT → Qwen3.5-4B → Qwen3-TTS — one command to start, no API key or cloud service needed.

Original post →

More from coding & agent

coding & agent channel →