Hugging Face Shows How to Run Reachy Mini's Voice Stack Fully Locally
Hugging Face
- Tutorials & guides
- Developer tools
- Robotics
- #reachy mini
- #speech-to-speech
- #vllm
Browse Tutorials & guides topic page →
A Hugging Face blog post explains how to run the Reachy Mini conversation app without sending audio to a server. It uses the speech-to-speech library, a cascaded VAD-STT-LLM-TTS pipeline exposed through a Realtime API-compatible WebSocket, with recommended open components and swappable stages.
What the publisher changed
Hugging Face published a walkthrough for running Reachy Mini's conversation app without a hosted speech server. Previously, the author says, audio had to be sent to a server. The new path runs the entire loop locally using the publisher's speech-to-speech library, described as a cascaded VAD → STT → LLM → TTS pipeline that exposes a Realtime API-compatible /v1/realtime WebSocket. Once the backend is launched, the robot is pointed at it from the conversation app's UI by editing the connection to select local mode. The author recommends llama.cpp serving Gemma 4 for the LLM, Silero VAD v5, Parakeet-TDT 0.6B v3 for STT and Qwen3-TTS, while stressing that cascades let each component be swapped. The LLM can also run in-process via MLX or Transformers, or sit behind a separate Responses API endpoint, and the engine can live on a laptop with the robot connecting over the local network.
A realistic scenario: an on-site robotics demonstrator
An engineer preparing a Reachy Mini for a demo where network access is restricted could try the local configuration so audio never leaves the machine they control, a privacy property the publisher cites as a main motivation. Because the pipeline is modular, they could start with the publisher's recommended components and test whether swapping in a different STT model suits their language needs — the author notes the defaults were optimized for multilingual use, which may not match a single-language deployment. A lab technician could similarly run the engine on a laptop and connect the robot over the LAN by binding to a LAN address rather than localhost, as the post describes.
How to trial it and measure success
A team could begin with the publisher's quick-start sequence: serve the LLM with llama.cpp, install the speech-to-speech library, and first validate conversation in local terminal mode before serving the robot. They should record a baseline — for example, the round-trip delay and failure rate they see with their current setup — then repeat identical scripted utterances and compare. Useful checks include whether speech is recognized correctly, whether responses start within an acceptable delay, and whether connections drop. The author flags that LLM inference is the main latency bottleneck, so testing with thinking disabled for conversational models, or enabling multi-token prediction where supported, are concrete variables to isolate. A human should review transcripts for accuracy and watch for tool-call failures; the post warns older vLLM versions trip when the assistant calls a tool.
Trade-offs, prerequisites and unknowns
Running locally trades convenience for control: the publisher frames it as gaining privacy, no per-minute or per-token API fees, and full component choice, but it requires suitable hardware and setup effort. The author is explicit that each stage involves trade-offs between speed, quality and language coverage, and that their defaults are opinions that can be swapped. The vLLM path requires version 0.21.0 or later for tool-call streaming, and the post recommends keeping thinking output off for natural conversation because each reasoning token adds silence the user hears. A team should not assume local mode is right if they lack the hardware, need frontier-model capability, or cannot maintain model updates; the post also notes that a separate Responses API mode can point at hosted providers. Benchmarks, supported hardware ranges and long-run reliability are not established in the supplied excerpt.
Context and limitations
- Drawn from a short excerpt of the publisher's blog post; the full article was not analyzed.
- Recommended models, versions and setup steps are the publisher's opinions and may change as components are updated.
- No independent performance measurements or latency benchmarks are presented in the supplied text.
- Based on bounded source text, not independent testing. Practical suggestions are editorial proposals.