Local Neural TTS for Reachy Mini
On-Device Inference
Runs entirely on the Raspberry Pi CM4 using ONNX Runtime — no cloud APIs, no external servers, complete privacy.
Audio-Reactive Movement
The robot's head moves naturally while speaking, with sway and roll synchronized to the audio output.
Performance
Generates speech at ~0.2x real-time on CM4 (2 seconds of audio takes ~10 seconds to generate). Optimized for embedded ARM hardware.
LSD Steps Parameter
Control the quality/speed tradeoff via the settings panel. Fewer steps = faster generation but lower quality. Default is 3 steps for best CM4 performance.