Open-source voice agent stacks Gemma 4, Parakeet, and Qwen3TTS on 9,000 Reachy robots
Hugging Face and Cerebras built an open-source real-time voice agent combining Gemma 4 VLM, Nvidia Parakeet ASR, and Alibaba Qwen3TTS, now deployed on 9,000+ Reachy Mini robots with sub-second latency.
Hugging Face and Cerebras released a working open-source real-time voice dialogue agent this week, combining three open models into a low-latency pipeline. The stack pairs Cerebras's fast inference for Gemma 4 VLM with Nvidia's Parakeet speech recognition and Alibaba's Qwen3TTS synthesis. The demo shows near-instant response times, and the system is already deployed on more than 9,000 Reachy Mini robots in the field.
The architecture routes spoken input through Parakeet for transcription, feeds the text to Gemma 4 VLM running on Cerebras hardware for dialogue generation, then synthesizes the reply with Qwen3TTS. Cerebras claims the inference speed on Gemma 4 keeps end-to-end latency minimal—users hear a response within a second of finishing a sentence. The GitHub repository includes the full pipeline code, and a live demo lets anyone test the agent in a browser.
What stands out
- 01Gemma 4 VLM on Cerebras — vision-language model inference accelerated by Cerebras hardware, handling dialogue context and multimodal input.
- 02Parakeet ASR from Nvidia — automatic speech recognition transcribing user audio into text with low word-error rates.
- 03Qwen3TTS from Alibaba — text-to-speech synthesis generating natural-sounding voice replies in real time.
- 04Reachy Mini deployment — the pipeline runs on over 9,000 Reachy Mini humanoid robots, demonstrating real-world scale beyond a research prototype.
- 05Open-source components — all three models and the orchestration code are publicly available, letting developers swap models or retrain for custom voices and languages.
