Hugging Face and Cerebras have demonstrated a real-time voice AI system that eliminates the frustrating delays typically found in speech-to-speech interactions.

The collaboration pairs Google DeepMind's Gemma 4 31B model with Cerebras' high-speed inference infrastructure to create conversations that flow with human-like responsiveness.

The system uses a modular, open-source architecture that chains together multiple components. Speech input flows through Nvidia's Parakeet for recognition, then to Gemma 4 inference on Cerebras hardware, before generating spoken responses via Alibaba's Qwen3TTS.

Cerebras addresses what the companies identify as the critical bottleneck: language model response time. While many production systems achieve reasonable median latency, they still suffer multi-second delays at the 95th percentile that make conversations feel unreliable.

"For voice AI, latency is a critical parameter," the companies said in their announcement. "Instead of waiting for an AI to respond, conversations flow with the responsiveness users expect from human interaction."

The technology already powers more than 9,000 Reachy Mini robots in real-world deployments. For robotics and embodied AI applications, the companies argue that responsiveness transforms interactions from feeling mechanical to feeling alive.

Why speed matters for voice AI

The collaboration reflects a broader industry challenge where model quality has advanced rapidly but user experience remains constrained by response times. Tool calls and multimodal processing can compound delays across multiple conversation turns.

Cerebras' inference speed provides stability at the performance tail, where occasional slow responses can undermine user confidence in voice systems.

The companies have made the complete pipeline available as an open-source project, allowing developers to inspect, modify, and extend each component for different applications.

The demonstration is accessible through Hugging Face Spaces, with the full codebase published in the huggingface/speech-to-speech repository for developer experimentation.