Andi from Hugging Face demonstrates a fully open-source voice pipeline using Gemma 4 31B

Andi from Hugging Face has released a fully open-source and free-to-use demo that creates a voice interaction pipeline. The system integrates Nvidia's parakeet, the Gemma 4 31B model served by Cerebras, and custom inference for Qwen3TTS.

• The stack functions as a drop-in replacement for OpenAI's realtime API.
• It is designed to see and search the web with low latency.
• Local execution is supported, with similar latencies achieved on a MacBook Pro M3 36GB using Gemma 4 E4B.
• A cloud-based web demo is available at hf-realtime-voice on Hugging Face Spaces.

This pipeline enables users to run local voice interactions and serves as the underlying technology for Reachy Minis.