vimal2023/astra-duplex-agent

Astra Duplex Agent is a voice AI assistant built for a hackathon, designed to solve a common frustration with talking to AI: most voice agents make you wait for them to finish speaking before you can interrupt, and if you do interrupt them, they forget what they were saying. This project lets a person talk to the AI and interrupt it mid-sentence, with the AI keeping track of exactly what it had already said before being cut off, so the conversation can continue naturally instead of restarting. The system listens for a person's voice using a sound-detection component called Silero VAD, which reacts in well under 50 milliseconds and immediately stops the AI's text generation and audio playback. The part of its reply that had already been spoken is saved to a database, so the next response can pick up from that exact point or shift smoothly to whatever new topic the person raised. Behind the scenes, the project uses a browser-based microphone and audio player connected over a live two-way connection to a Python server. That server converts speech to text using Groq's hosted version of Whisper, generates responses using a Llama language model, and converts the reply back to speech using ElevenLabs. A tool called LangGraph coordinates the steps of listening, thinking, and speaking as a repeatable cycle. The project is written mostly in Python for the server and React for the browser interface, and it is released under the MIT license, meaning it can be freely used, copied, and modified, including for commercial purposes. A recorded video demo is linked in the project description for anyone who wants to see it working before trying it themselves.