Hugging Face and Cerebras bring Gemma 4 to real-time voice AI
Hugging Face and Cerebras unveiled a real-time speech-to-speech AI pipeline using Gemma 4 31B, reducing latency for natural voice interactions in robots and assistants.
Hugging Face and Cerebras announced a new open, modular speech-to-speech architecture designed to improve real-time voice AI interactions. The system integrates Google DeepMind’s Gemma 4 31B language model, Qwen for text-to-speech, and Cerebras’ high-speed inference to minimize response delays. Each component remains open and replaceable, allowing developers to customize the stack for various applications. The collaboration aims to address persistent latency issues that disrupt conversational flow in current voice AI systems.
Cerebras’ role focuses on accelerating language model inference, a major bottleneck in voice AI pipelines. Traditional systems often achieve acceptable median response times but suffer from unpredictable delays at higher percentiles, such as P95. By reducing these delays, Cerebras enhances the reliability of real-time interactions, particularly in edge cases. The architecture targets embodied AI, robots, and assistants where responsiveness directly impacts user experience. The demo highlights how faster inference enables smoother, more natural conversations compared to existing solutions.
The pipeline is already deployed in over 9,000 Reachy Mini robots, demonstrating its practical application in real-world scenarios. For these robots, low latency is not merely an enhancement but a necessity to create lifelike interactions. The open design allows developers to modify or extend components, fostering innovation across research and commercial projects. This approach contrasts with proprietary, closed systems that limit flexibility and scalability.
The partnership underscores a commitment to open-source AI and high-performance inference, positioning real-time voice AI as the next frontier. Developers are encouraged to test the demo, access the code, and contribute to advancing conversational AI. The initiative reflects a broader trend toward combining open models, infrastructure, and breakthrough speed to redefine user experiences in voice-driven applications.