source: hugging face blog: hugging face and cerebras bring gemma 4 to real-time voice ai
level: technical
hugging face and cerebras have built a real-time voice ai demo that pairs open models with fast inference. the system uses a cascaded speech-to-speech stack: nvidia's parakeet for speech recognition, google deepmind's gemma 4 31b as the language model running on cerebras hardware, and alibaba's qwen3tts for text-to-speech. each component is modular and open, so developers can swap parts for different assistants, robots, or research projects.
latency is the main problem this setup addresses. many voice ai systems have acceptable median response times but still suffer from occasional multi-second delays, especially at the 95th percentile. those slow responses make conversations feel unreliable. cerebras speeds up the language model step, which is often the bottleneck, and makes performance more stable even under heavy load. this means the whole pipeline can deliver consistently fast replies.
the same pipeline already runs on reachy mini robots, with over 9,000 units in the field. for robots and embodied ai, responsiveness is not just a nice feature—it makes interactions feel alive. the collaboration shows that open-source models and open infrastructure can achieve the low latency needed for natural, real-time conversations at scale. developers can try the demo on a hugging face space and find the code in the huggingface/speech-to-speech repository.
why it matters: faster, more stable language model inference reduces awkward delays in voice ai, making open-source stacks practical for real-world assistants and robots.
source: hugging face blog: hugging face and cerebras bring gemma 4 to real-time voice ai