license: mit
task_categories:

• audio-classification
• audio-to-audio
  tags:
• audio
• music
• source-separation
• musdb18
• stems
• active-segments
• cs229
• stanford
  pretty_name: "MUSDB18 Active Stems - CS229 Project"
  size_categories:
• 10K<n<100K

MUSDB18 Active Stems Dataset - CS229 Project

This dataset contains active stem segments extracted from the MUSDB18 dataset for the Stanford CS229 Machine Learning course project on audio source separation.

Dataset Description

This is a processed version of the MUSDB18 dataset containing only the active segments of each stem (drums, bass, vocals, accompaniment, and mixture), designed to improve training efficiency for music source separation models.

Key Features

• Active Segment Detection: Only segments where stems have significant energy
• 5 Stems: mixture, drums, bass, vocals, accompaniment
• Consistent Format: 22.05 kHz sample rate, mono audio
• Rich Metadata: Detailed segment information and statistics

CS229 Project Context

This dataset was created as part of a Stanford CS229 course project focusing on:

• Music source separation using deep learning
• Comparison of different neural architectures (Conv-TasNet, etc.)
• Analysis of active vs. inactive audio segments in training

Dataset Structure

extracted_stems/
├── train/              # Training split
│   ├── drums/         # Active drum segments
│   ├── bass/          # Active bass segments  
│   ├── vocals/        # Active vocal segments
│   ├── accompaniment/ # Active accompaniment segments
│   └── mixture/       # Active mixture segments
├── test/              # Test split (same structure)
└── metadata/          # JSON metadata files

Quick Start

Loading with datasets library

from datasets import load_dataset

Load the full dataset
dataset = load_dataset("cs229-audio-ml-project/musdb18-processed")
Access training data
train_data = dataset["train"]
for item in train_data:
    audio = item["audio"]["array"]
    stem_type = item["stem_type"]
    track_name = item["track_name"]

Manual loading

import soundfile as sf
import json

Load an audio segment
audio, sr = sf.read("train/vocals/track_vocals_001.wav")
Load metadata
with open("metadata/train_metadata.json") as f:
    metadata = json.load(f)

Extraction Parameters

• Segment Length: 4.0 seconds
• Hop Length: 2.0 seconds (50% overlap)
• Energy Threshold: 0.01 RMS
• Sample Rate: 22,050 Hz
• Minimum Duration: 1.0 seconds

Citation

If you use this dataset in your research, please cite:

@dataset{cs229_musdb18_active_stems,
  title={MUSDB18 Active Stems Dataset},
  author={CS229 Audio ML Project Team},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/cs229-audio-ml-project/musdb18-processed}
}

Original MUSDB18 citation:

@misc{musdb18,
  author       = {Rafii, Zafar and Liutkus, Antoine and Stöter, Fabian-Robert and Mimilakis, Stylianos Ioannis and Bittner, Rachel},
  title        = {MUSDB18-HQ - an uncompressed version of MUSDB18},
  month        = {December},
  year         = {2019},
  doi          = {10.5281/zenodo.3338373},
  url          = {https://doi.org/10.5281/zenodo.3338373}
}

Contact

For questions about this dataset or the CS229 project, please open an issue in this repository.

Created for Stanford CS229 - Machine Learning Course Project