Hugging Face and Cerebras demo real-time speech-to-speech pipeline using Gemma 4 31B

Hugging Face and Cerebras publicly demonstrated a real-time speech-to-speech pipeline that integrates Google DeepMind’s Gemma 4 31B vision-language model with Cerebras inference hardware and Alibaba’s Qwen3TTS text-to-speech engine. The architecture is designed as an open, cascaded stack where each stage—speech recognition, language-model inference, and speech synthesis—can be inspected, modified, or replaced by developers.

The partners highlight that many production voice systems achieve acceptable median latency but still suffer multi-second delays at the P95 percentile, which undermines user experience. By accelerating language-model inference with Cerebras hardware, the pipeline aims to reduce those long-tail delays and deliver more predictable, sub-second response times across turns.

The demonstration builds on a Hugging Face speech-to-speech pipeline already deployed on more than 9,000 Reachy Mini robots, underscoring the importance of responsiveness for embodied AI. The collaborators argue that low latency and predictable performance are not merely cost-saving measures but prerequisites for interactions that feel natural at scale.

To enable experimentation, the teams published a Hugging Face Space and an open repository containing the full pipeline code. Developers can run the live demo and inspect or extend the modular components for their own assistants, robots, or research projects.