Hugging Face and Cerebras bring Gemma 4 to real-time voice AI
Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated.
A fully open cascaded speech-to-speech stack—Parakeet STT → Gemma 4 31B on Cerebras → Qwen3TTS—now runs fast enough to hit conversational latency, with the key win being tail stability (P95) rather than just median, which is what actually kills voice UX. If you're building voice agents or embodied AI, this is a swappable, self-hostable alternative to closed real-time APIs, and the same pipeline is already in production on 9,000+ Reachy Mini robots. The bet is that Cerebras's inference speed makes multi-turn tool-calling and multimodal steps feel real-time, so the constraint you're designing around shifts from LLM response time to your STT/TTS choices.
Gemma 4 31B is now shown in an open speech-to-speech loop on Cerebras, with Parakeet ASR and Qwen3TTS, targeting low and stable latency rather than just better median response time. For production voice agents and robots, the key implication is that the LLM step can stop being the long-tail bottleneck, making natural turn-taking feasible in modular open stacks already deployed across 9,000+ Reachy Mini robots.