deezer/spleeter
Spleeter
Spleeter is an AI audio source separation library and deep learning toolkit designed to split mixed music files into individual audio stems, such as vocals and drums. It provides a suite of pretrained models for isolating different instruments and voices from a recording.
The toolkit includes capabilities for training and evaluating custom audio separation models using labeled datasets and configuration files. It also features utilities for measuring model performance by comparing separation outputs against reference datasets.
The system manages audio processing through spectral representations and uses a custom interface for loading and saving audio data across different storage formats. Exporting separated stems is handled via asynchronous processing.
Features
- Audio Stem Extractors - Splits mixed music files into individual audio tracks like vocals and drums using deep learning models.
- Audio Source Separation Models - Splits mixed audio files into individual stems like vocals and drums using pretrained machine learning models.
- Separation Model Training - Implements capabilities for creating custom audio separation models using labeled datasets.
- Spectral Masking - Uses spectral masking to filter out unwanted frequencies and isolate individual audio stems from a mixture.
- Pretrained Model Integrations - Enables the application of pretrained deep learning model weights to new audio inputs for source separation.
- Deep Learning Audio Libraries - Functions as a deep learning library for splitting music files into individual audio stems.
- Deep Learning Toolkits - Provides a comprehensive toolkit for training, evaluating, and exporting custom audio separation models.
- Source Separation Training - Provides the ability to train custom audio separation models using labeled datasets and configuration files.
- U-Net Architectures - Employs a U-Net encoder-decoder architecture to learn spectral patterns for isolating audio sources.
- FFT Spectral Processing - Converts audio waveforms into time-frequency representations using Short-Time Fourier Transforms for processing and reconstruction.
- Model Performance Evaluators - Quantifies the accuracy and quality of source separation models by comparing outputs against reference datasets.
- Training Configurations - Provides external configuration files to manage hyperparameters and dataset paths for automated model training.
- AI and Machine Learning - Audio source separation library.
- Natural Language Processing - Source separation library for audio.
- Audio Processing - Source separation library for isolating audio tracks.
