music-interpolation

Mix between music tracks using machine learning

Introduction

This repository is a sandbox to experiment with machine learning techniques for smoothly transitioning between music tracks.

Development

1. git clone https://github.com/jhurliman/music-interpolation.git
2. cd music-interpolation
3. Install pipenv if you have not already: pip install pipenv
4. Install dependencies: pipenv sync --python 3.10 --dev
5. Run tests: pipenv run pytest

Usage

import librosa
from IPython.display import Audio
from music_interpolation import EncodecInterpolation

# Instantiate the interpolation class, fetching and loading the pre-trained
# Encodec model
interp = EncodecInterpolation()

# Load two audio tracks, resampling to the sampling rate of the Encodec model
# (defaults to 48kHz)
audio_a, _ = librosa.load('audio_a.mp3', sr=interp.sampling_rate, mono=False)
audio_b, _ = librosa.load('audio_b.mp3', sr=interp.sampling_rate, mono=False)

# Transition between two audio tracks using linear interpolation between
# embedding vectors generated by a pre-trained Encodec model
audio_c = interp.interpolate(audio_a, audio_b)
Audio(data=audio_c, rate=interp.sampling_rate)

Contributing

Contributions are welcome! Open a pull request to fix a bug, or open an issue to discuss a new feature or change.

License

This project is licensed under the MIT License - see the LICENSE file for details.