Hugging Face and Cerebras bring Gemma 4 to real-time voice AI
Hugging Face
Hugging Face and Cerebras just built a voice AI demo that talks back almost instantly, using Gemma 4 and Qwen's TTS. It's fully open-source, so anyone can pull it apart and rebuild it for their own robot or app.
Based on reporting by Hugging Face — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Latency has always been the awkward silence in voice AI. You ask a question, the AI understands you fine, but then there's that beat — sometimes two, three seconds — before it answers. Hugging Face and Cerebras think they've found a way to close that gap without sacrificing the openness that makes these systems hackable in the first place.
The setup is a cascaded speech-to-speech pipeline, meaning it's built from separate, swappable pieces rather than one monolithic model. Audio comes in through Nvidia's Parakeet for speech recognition, gets handed to Google DeepMind's Gemma 4 31B running on Cerebras hardware for the actual language reasoning, and comes back out through Alibaba's Qwen3TTS for the spoken response. Every layer is open source, so a developer building a robot, a customer-service bot, or a research prototype can swap in a different speech recognizer or a different voice without asking anyone's permission.
The real story here is what Cerebras brings to the table: raw inference speed on the Gemma 4 layer, which happens to be the part of the pipeline most responsible for those maddening delays. Hugging Face points out that plenty of production voice systems already have decent median response times — the problem is the tail end, the P95 cases where a tool call or a multimodal step turns a snappy conversation into a stuttering one. Cerebras's pitch isn't just
My take — AI-written commentary, not fact-checked reporting
I'll believe the 'feels alive' claim when I talk to it myself, but pairing an open model stack with dedicated inference hardware instead of a black-box API is exactly the kind of move that keeps voice AI from becoming another walled garden — more of this, please, and less proprietary latency theater.
Read more about this at: Hugging Face