Hugging Face and Cerebras wired Google’s Gemma 4 into a full speech-to-speech pipeline that runs at conversational latency — entirely on open-source components.
The stack chains three stages into a single real-time loop:
- Speech-to-text — Whisper transcribes audio input on the fly
- Reasoning — Gemma 4 processes the transcript on Cerebras inference hardware, which pushes over 2,200 tokens per second on comparable models like Llama 3.3 70B
- Text-to-speech — A TTS model streams audio back before the full response is even generated
The result is voice AI that feels like a phone call, not a walkie-talkie. Each stage hands off to the next with streaming, so the user hears the first words of a reply while the model is still generating the rest. Cerebras’s wafer-scale chips handle the middle step — the actual thinking — fast enough that it stops being the bottleneck.
What makes this worth watching: every piece is open-source and available on the Hugging Face Hub. Developers can swap components — different STT model, different TTS voice, different LLM — without rebuilding the pipeline. That’s a direct contrast to closed voice APIs from OpenAI or Google, where you get one black box and no control over latency tradeoffs.
NOTE
Cerebras inference has been available on Hugging Face since early 2025 via API. This voice pipeline is a reference architecture built on top of that integration, not a new product.
The practical upshot: if you’re building voice features into an app, this is a working open-source template you can deploy now. Gemma 4 handles the reasoning, Cerebras handles the speed, and you own the whole stack.