We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
Hugging Face Blog·May 27·5 min read
Reachy Mini goes fully local
TL;DR
Running AI assistants on robots previously required cloud services and API keys, creating privacy and latency concerns. Hugging Face demonstrated a fully local speech-to-speech pipeline using llama.cpp that chains voice activity detection, speech-to-text, language model, and text-to-speech components.
✦ Why It Matters
Engineers can now deploy private, low-latency conversational robots without cloud infrastructure or API costs.
Key Takeaways
- Install the speech-to-speech library using pip.
- Run llama.cpp to serve the language model locally.
- Configure the conversation app to connect to your local server.
- Explore different models for VAD, STT, and TTS to optimize performance.
How It Works
The local speech-to-speech pipeline consists of four stages: Voice Activity Detection (VAD) identifies when speech occurs, Speech-to-Text (STT) converts spoken words into text, a Language Model (LLM) processes the text to generate responses, and Text-to-Speech (TTS) synthesizes spoken responses. By using a cascade approach, users can swap components for better performance or quality, allowing for a customizable and efficient conversational experience.