RealtimeVoiceChat lets you hold a natural, spoken conversation with a large language model, hearing answers read back to you in near real time. Instead of typing prompts, you simply talk, and the system streams your voice to a speech pipeline that transcribes it, sends the text to an LLM, and synthesizes the reply as audio with minimal delay.
RealtimeVoiceChat is an open-source project on GitHub with about 3,860 stars, most recently updated in 2025. It combines the author’s RealtimeSTT transcription engine and RealtimeTTS synthesis library into a client-server application built for low-latency interaction. Your voice is captured in the browser, streamed over WebSockets to a Python backend, transcribed, answered by a model such as a local Ollama model or OpenAI, and returned as spoken audio. The original author has stepped back from active development, so the project is now community-driven, with pull requests still reviewed and merged. Because the transcription and synthesis stages are modular, you can swap voices, languages and models to suit your setup, and the whole stack runs on a normal PC with a microphone.
RealtimeVoiceChat is free and open source. You can clone the repository and run it yourself; the only costs are whatever you pay for the LLM and speech services you choose to connect, and fully local setups with Ollama avoid those entirely.
Developers and hobbyists who want a hands-free voice interface for an LLM, whether for accessibility, prototyping voice assistants, or experimenting with real-time speech pipelines on their own machine.