Recho

Recho is a fully local AI chatbot with voice support. It implements a speech-to-speech pipeline based on open-source models like Whisper (speech to text), Supertonic (speech to text) and a local language model at its core.

Specifically, it uses whisper.cpp via whisper-rs and a custom Supertonic implementation that uses onnxruntime (ort) under-the-hood.

I built this project primarily to improve my Rust skills. This isn't a robust application, nothing fancy, but it works.

Demo

recho_v2_demo.webm

Building and Running

Requirements

• Rust (stable) Install via rustup
• Any LLM inference engine with OpenAI-like API (like Ollama, llama.cpp, etc.)

Building

Clone the repository and build the project using Cargo:

cargo build

Usage

Pre-requisites

• An LLM running via inference engines like Ollama, llama.cpp, etc.
• A Whisper model (like whisper-ggml-small.bin)
• A Supertoic TTS ONNX models and voice styles

Run the project like this:

cargo run -- -l "http://localhost:1234/v1/chat/completions" -o "supertonic/models" -v "supertonic/voices/M1.json" -w "whisper/ggml-small.bin"

Tested on a Linux system with LMStudio for serving LLMs.

Credits

• thewh1teagle's piper-rs
• supertone-inc's supertonic