Hugging Face and Cerebras Join Forces for Real-Time Voice AI Based on Gemma 4

Hugging Face, together with Cerebras, has created an open-source real-time speech synthesis pipeline that makes voice interaction with AI as natural as possible. The system operates on the principle of "speech input -> speech recognition via Nvidia Parakeet -> output from the multimodal language model Gemma 4 (developed by Google DeepMind) on Cerebras accelerators -> speech synthesis via Alibaba's Qwen3TTS -> speech output." Each component is open, modular, and replaceable, allowing developers to adapt the stack for their assistants, robots, or research projects. Particular attention is paid to latency stability: although many systems show acceptable median latency, at P95 (the 5% slowest requests), multi-second pauses still occur. Cerebras solves this problem by ensuring fast and predictable language model inference. This same pipeline is already being used in Reachy Mini robots, with over 10,000 units in existence. Developers are invited to try the demo on Hugging Face Space and experiment with the code from the huggingface/speech-to-speech repository.