Hugging Face and Cerebras demo real-time voice AI with low latency

What happened

Hugging Face and Cerebras demonstrated a speech-to-speech AI pipeline that combines open-source models—Nvidia's Parakeet for speech recognition, Google DeepMind's Gemma 4 31B language model running on Cerebras inference, and Alibaba's Qwen3TTS for text-to-speech—to enable natural, fast conversational responses. The modular, open architecture allows developers to inspect, modify, and extend each component.

Why it matters

Latency has been a major bottleneck in voice AI systems; many production systems experience multi-second delays at the P95 (worst-case scenarios), making conversations feel unreliable and unnatural. By making language-model inference dramatically faster and more stable, Cerebras addresses this bottleneck. For robots, voice assistants, and embodied AI, this responsiveness is not a cosmetic improvement—it makes interactions feel alive and natural at scale, which may enable more practical deployment of conversational AI in real-world applications.

What to watch

The pipeline already powers Reachy Mini robots, with more than 9,000 robots in the wild. Developers can explore the demo on Hugging Face Space and the code in the huggingface/speech-to-speech repository.