Hugging Face and Cerebras bring Gemma 4 to real-time voice AI
02:00 · July 1, 2026 · Hugging Face Blog

Summary
Hugging Face and Cerebras have developed an open, modular speech-to-speech pipeline that combines several components to reduce response delays in voice interactions. The system uses a cascaded architecture with automatic speech recognition feeding into a large language model and then into text-to-speech synthesis. Nvidia’s Parakeet handles the initial transcription step, Gemma 4 31B runs on Cerebras hardware for inference, and Alibaba’s Qwen3TTS generates the final audio output.
The design specifically addresses P95 latency, the tail-end delays that remain noticeable even when median response times appear acceptable. Cerebras inference improves both speed and predictability at this percentile, which becomes critical when tool calls or additional processing steps extend the chain. This stability supports more fluid exchanges than systems that deliver occasional multi-second pauses.
Each stage of the pipeline remains independently replaceable, allowing developers to substitute models or adapt the stack for specific robots, voice assistants, or research prototypes. The same pipeline already operates on more than 9,000 Reachy Mini robots, where low and consistent latency determines whether spoken interaction feels responsive rather than mechanical. The collaboration emphasizes open components and infrastructure that can be inspected or extended without proprietary restrictions.
Why it matters
This article is highly relevant for ML Engineers as it provides a practical, open-source architecture for solving critical latency bottlenecks in real-time voice AI. Dutch AI teams can directly implement this modular stack using the provided repositories to build responsive conversational agents and embodied AI solutions.








