Musical Dataset Builder

This project provides a powerful and automated workflow for creating clean, single-instrument datasets from a collection of mixed audio tracks. By leveraging the Lalal.ai API for high-quality source separation and sox for audio manipulation, this tool streamlines the process of isolating specific instruments from complex musical pieces.

Features

• Recursive Audio Discovery: Automatically finds all audio files in a specified directory.
• Silence Removal: Uses a two-pass sox process to remove silence from the beginning, middle, and end of audio files.
• Audio Segmentation: Splits long audio files into smaller, manageable segments.
• Source Separation: Integrates with the Lalal.ai API to extract specific instrumental stems (e.g., wind, strings, piano).
• Clear Logging: Provides detailed logs for easy debugging and progress tracking.

Setup

1. Create and activate a virtual environment:

   python -m venv venv
   venv\Scripts\activate

2. Install the required dependencies:

   pip install -r requirements.txt

Usage

Run the script from your terminal with the following command, providing the path to your dataset, the instruments you want to extract, and your Lalal.ai API key.

python process_audio.py --dataset_path /path/to/your/audio/files --instruments winds strings --api_key YOUR_LALALAI_API_KEY

Flags

--dataset_path: Path to the dataset folder.

--segment_duration: Duration of audio segments in minutes.

--instruments: List of instruments to extract (e.g., winds, strings). The supported stems are: wind, strings, piano, bass, electric_guitar, acoustic_guitar, synthesizer, synth, drum, voice, accompaniment, ensemble, percussion, and effects.

--api_key: Lalal.ai API key.

--output_path: Path to save the processed files. If not provided, the processed files will be saved in the same folder as the input files.

--no_stem_separation: Skip the upload to Lalal.ai. If provided, the script will not send the audio files to Lalal.ai for source separation.