Hugging Face and Cerebras Build Real-Time Voice AI with Gemma 4

Voice AI has made impressive progress in model quality, but real-world user experience remains constrained by latency.

Many production systems achieve acceptable median response times while still suffering from multi-second delays at the P95, making conversations feel unreliable and unnatural.

The tension is between increasingly capable models and the unpredictable, slow inference that breaks the illusion of real-time interaction—especially when tool calls or multimodal steps require multiple turns.