Auracle

Auracle is a semantic audio search tool built with Tauri and React. It allows you to index your local audio files and search through them using natural language descriptions or files - powered by CLAP (Contrastive Language-Audio Pretraining).

Screen.Recording.2026-04-28.at.15.30.31.mov

✨ Features

• Semantic Audio Search: Search your local audio library using text queries like "electronic drum beat" or "bird sounds".
• Local Machine Learning: Runs CLAP entirely locally using ONNX Runtime. No internet connection required, and your audio never leaves your machine.
• Fast Vector Search: Uses sqlite-vec for instant, local vector similarity search directly within SQLite.
• Modern UI: Built with React, TailwindCSS, and Radix UI.
• Cross-Platform: Built on top of Tauri for native performance on Windows, macOS, and Linux.

🛠️ Tech Stack

Frontend

• React 19 & Vite
• TailwindCSS 4
• Radix UI primitives and Lucide Icons
• TypeScript

Backend

• Tauri 2 (Rust)
• SQLite (via sqlx) with sqlite-vec for vector embeddings syntax.
• ONNX Runtime (ort) for running the machine learning models.
• Symphonia for audio decoding.
• rubato for audio resampling.
• tokenizers (HuggingFace) for text embeddings parsing.

🚀 Getting Started

Prerequisites

• Node.js (v18+)
• Rust (Latest stable)
• Tauri prerequisites (depends on your OS)

Installation

1. Clone the repository:

   git clone https://github.com/ruarim/auracle.git
   cd auracle

2. Install frontend dependencies:

   npm install

3. Extract CLAP ONNX Models:

   • run ./src-py/export-clap.py and move the .onnx and .onnx.data files to ./src-tauri/models

4. Run the development server:

   npm run tauri dev

🏗️ Architecture

Auracle's core functionality relies on processing audio files into Mel Spectrograms and embedding them into a high-dimensional vector space using the CLAP audio model.

1. Indexing: When a file is added, it is decoded and resampled with symphonia and rubato, respectively, convolution of the STFT (realfft) with a filter matrix results in a Mel-spectrogram. The ONNX audio model generates an embedding, which is then stored in SQLite utilizing sqlite-vec.
2. Searching: Text queries are tokenized and passed through the ONNX text model. The resulting text embedding is queried against the sqlite-vec index to find the closest matching audio files based on cosine distance.

🎯 To-Do

• Advanced search filters
• Retry failed audio analysis
• Drag and drop folder/file support
• Support and display all audio file types and bit-depths
• Confirmation dialog on "Delete All"
• Comprehensive error handling

📝 License

MIT License