Hugging Face and Cerebras launch an open real-time speech-to-speech AI pipeline on Gemma 4

Hugging Face and Cerebras have released an open, cascaded speech-to-speech voice AI demo built on Google DeepMind's Gemma 4 31B, pairing Cerebras's inference speed with a fully open, swappable model stack to close the latency gap that makes most voice assistants feel unnatural.

What's new

The pipeline chains three open components into a single real-time loop: speech recognition via Nvidia's Parakeet, language understanding and generation via Gemma 4 31B running on Cerebras, and text-to-speech via Alibaba's Qwen3TTS. Hugging Face and Cerebras describe it as "the strength of the open-source AI ecosystem" assembled into one stack, with every layer inspectable, modifiable, and replaceable by developers building their own assistants, robots, or products.

The announcement frames the core problem as tail latency, not average latency: many production voice systems post a reasonable median response time but still hit frustrating multi-second delays at the P95, especially once tool calls or multimodal steps are involved. Cerebras's role is to remove the language-model response time as a bottleneck so the rest of the pipeline — recognition and speech synthesis — isn't waiting on it.

The same underlying Hugging Face speech-to-speech pipeline already powers Reachy Mini robots, with more than 9,000 units reportedly in the wild, giving the release a real deployment base rather than just a demo. A live version is published as a Hugging Face Space (smolagents/hf-realtime-voice), with the code open-sourced under huggingface/speech-to-speech on GitHub.

Context

The release lands alongside a wave of voice-AI activity across the industry — ElevenLabs, OpenAI's realtime API, and Google's own Gemini Live have all pushed low-latency conversational voice this year. Hugging Face's angle is openness: rather than a closed API, this is a fully swappable open pipeline that any developer can inspect and redeploy, with Cerebras supplying the inference layer that was previously the industry's main latency bottleneck.

Why it matters

Latency, not raw model quality, has been the main thing separating AI voice assistants from feeling conversational versus feeling like a phone tree. By open-sourcing a full cascaded stack — ASR, LLM, TTS — and pairing it with dramatically faster inference, Hugging Face and Cerebras are making it easier for smaller teams and robotics builders to ship voice interfaces that don't require negotiating access to a closed real-time API. The Reachy Mini deployment numbers suggest this isn't purely theoretical: it's already running at meaningful scale in physical robots, which is a harder bar than a chat demo.