Reachy Mini Now Runs Speech Conversations Entirely Local, No Cloud Required
Hugging Face enables fully on-device speech-to-speech for Reachy Mini robots using open-source models and cascaded AI pipelines.
Last verified:
Local-First Conversational Robotics Now Shipping
Reachy Mini operators can now deploy the entire conversational pipeline on their own hardware, according to a Hugging Face Blog post. The shift eliminates dependency on cloud services and ensures no audio data leaves the user’s system. The stack—powered by a cascaded voice activity detection → speech-to-text → LLM → text-to-speech pipeline—exposes a Realtime API-compatible WebSocket interface that the robot’s UI can point to once the backend launches.
Recommended Model Stack and Installation
Hugging Face recommends pairing Gemma 4 (served via llama.cpp) as the language model backbone with Silero for voice activity detection, Parakeet-TDT for speech-to-text, and Qwen3 for text-to-speech synthesis. The LLM setup uses a 65,536-token context window split across two parallel inference slots, flash attention for reduced latency and memory footprint, and full sliding-window attention caching to accelerate prompt processing on Gemma 4 specifically.
Installation is straightforward: brew install llama.cpp or winget install llama.cpp to get the server runtime, then launch via llama-server with flags for model selection, parallel slots, and attention optimizations. The speech-to-speech library installs via uv pip install speech-to-speech. On first run, the system downloads model weights; subsequent launches remain fast due to local caching.
Cascade Architecture Enables Model Swapping
The cascade approach—connecting distinct VAD, STT, LLM, and TTS components in series—is positioned as the most flexible design pattern in the open-source landscape. Because individual models can be substituted, users are not locked into a single vendor or version. Hugging Face notes that new models ship weekly, meaning operators can upgrade individual pipeline stages without redeploying the entire stack. This modularity contrasts with end-to-end speech models, which require full retraining to swap any component.
Why This Matters
For robotics developers and Reachy Mini users, local inference eliminates the latency penalty and privacy concerns of cloud APIs. Teams building conversational agents in regulated environments—healthcare, finance, or sensitive research—can now certify that no audio leaves their facility. The cascade architecture’s flexibility also means that as stronger open-weights models emerge (e.g., faster VAD, higher-accuracy STT, lower-latency TTS), users can cherry-pick improvements without waiting for a monolithic model release. If the recommended stack’s performance benchmarks hold up under production use, this pattern may influence broader adoption of modular inference pipelines in consumer robotics.
Frequently Asked Questions
Does Reachy Mini still require internet to run conversations?
No. The latest release enables fully local operation using llama.cpp and open-source speech models, with no API calls or data transmission to external servers.
Which models does the recommended stack use?
The stack uses Gemma 4 (LLM via llama.cpp), Silero for voice activity detection, Parakeet-TDT for speech-to-text, and Qwen3 for text-to-speech.
Can I swap out the recommended models?
Yes—the cascade architecture lets you substitute any compatible VAD, STT, LLM, or TTS model. New models are released frequently.
What hardware do I need?
The post uses Gemma 4 with a 64K context window and flash attention, which runs on modern consumer hardware; exact requirements depend on your model choices.