Show HN: Sonia. Local Voice Assistant Linux. 13Gb RAM, No GPU, Skills Supported

Category: ai-ml

Tags: voice-assistant, local-inference, full-duplex, tool-calling, typescript, llm, speech

Score: 7.8/10 (Innovation: 7, Technical: 9, Documentation: 9, Utility: 6)

Sonia is a local, full-duplex voice assistant built in TypeScript that combines streaming ASR, TTS, VAD, and an external OpenAI-compatible LLM endpoint to achieve barge-in support without a dedicated GPU, running under 15GB RAM. It features sophisticated turn detection, tool calling with a safety gate, bilingual support, and externally managed skills, demonstrating a complex real-time audio pipeline and thoughtful engineering trade-offs.

Target audience: AI/ML engineers and hobbyists interested in local, privacy-preserving voice assistants; backend developers exploring real-time audio pipelines

Repository: https://grigio.org/sonia-local-voice-assistant-in-less-than-13gb-ram-no-dedicated-gpu-skills-tool-calling-supported/ · TypeScript · GPL-3.0 · 57 stars

View on Hacker News