huggingface.coSourceSource Linked

Hugging Face and Cerebras bring Gemma 4 to real-time voice AI

Hugging Face demos real-time speech-to-speech from open parts: Nvidia Parakeet ASR, Gemma 4 31B on Cerebras inference, Alibaba's Qwen3TTS — pipeline code is open and already runs on 9,000+ Reachy Mini robots.

huggingface.co

Open Source Open Markdown Open JSON

Source Summary

**The gist** Hugging Face and Cerebras show a real-time voice pipeline built entirely from open components: Nvidia's **Parakeet** for speech recognition, **Gemma 4 31B** running on Cerebras inference, and Alibaba's **Qwen3TTS** for output. The code is in the **huggingface/speech-to-speech** repo with a Spaces demo, and the stack already ships on **9,000+ Reachy Mini** robots.

Practical Implication

**Why it matters** A modular ASR-to-LLM-to-TTS chain on a fast inference provider is now a copyable recipe for adding voice to an agent product without a proprietary end-to-end voice model — each stage swaps independently, so you can trade model quality against latency per stage.

Agent-Ready Context

**The gist** Hugging Face and Cerebras show a real-time voice pipeline built entirely from open components: Nvidia's **Parakeet** for speech recognition, **Gemma 4 31B** running on Cerebras inference, and Alibaba's **Qwen3TTS** for output. The code is in the **huggingface/speech-to-speech** repo with a Spaces demo, and the stack already ships on **9,000+ Reachy Mini** robots.

**Why it matters** A modular ASR-to-LLM-to-TTS chain on a fast inference provider is now a copyable recipe for adding voice to an agent product without a proprietary end-to-end voice model — each stage swaps independently, so you can trade model quality against latency per stage.

**Watch out** The post claims low latency but publishes **no benchmarks** — no tokens per second, no **P95 latency** figures — and no Cerebras pricing. Measure the pipeline on your own workload before depending on it.

Context Map

modelaudio#open-models

Uncertainty

The post claims low latency but publishes **no benchmarks** — no tokens per second, no **P95 latency** figures — and no Cerebras pricing. Measure the pipeline on your own workload before depending on it.