metadata

license: cc-by-4.0
task_categories:
  - audio-classification
  - audio-to-audio
tags:
  - audio
  - music
  - choral
  - satb
  - source-separation
  - pytorch
pretty_name: SATB Choral Source Separation Dataset (Compressed)
size_categories:
  - 1K<n<10K

SATB Choral Source Separation Dataset (Compressed)

This dataset contains preprocessed 4-second audio chunks from the Choral Singing Dataset (CSD) formatted for SATB (Soprano, Alto, Tenor, Bass) voice source separation tasks.

This is the compressed version with 8kHz sample rate and int8 precision for smaller file sizes.

Dataset Description

Total chunks: 26,883
Dataset size: ~4155.3 MB
Audio format: 8 kHz mono, 4-second chunks
Precision: int8 for memory efficiency
Sources: 4 individual voice parts (SATB)
Mixed audio: Sum of all 4 voices

Data Structure

Folder Structure

├── chunks/                    # All individual chunk .pt files
├── quality_samples/           # Sample WAV files for quality checking
├── train_split.pt            # List of training chunk filenames
├── val_split.pt              # List of validation chunk filenames
└── test_split.pt             # List of test chunk filenames

Chunk Files

Each .pt file in chunks/ contains a dictionary with:

mixed: Mixed audio tensor (4 seconds, 32k samples at 8kHz) - int8
sources: Individual voice tensors [4, 32k] (soprano, alto, tenor, bass) - int8
song: Song identifier ("ER", "LI", or "ND")
singer_combo: List of singer IDs used [soprano_id, alto_id, tenor_id, bass_id]
chunk_id: Chunk number within the combination

Usage

import torch

# Load data splits
train_files = torch.load('train_split.pt')
val_files = torch.load('val_split.pt')
test_files = torch.load('test_split.pt')

# Load a single chunk
chunk = torch.load(f'chunks/{train_files[0]}')
mixed_audio = chunk['mixed'].float() / 127.0  # Convert back to float
separated_voices = chunk['sources'].float() / 127.0

# mixed_audio shape: [32000] (4 seconds at 8kHz)
# separated_voices shape: [4, 32000] (SATB voices)

Compression Details

Sample rate: Reduced from 16kHz to 8kHz (2x reduction)
Precision: int8 instead of float16 (2x reduction)
Total compression: ~4x smaller than original
Quality samples: Available in quality_samples/ for evaluation

Original Dataset

Based on the Choral Singing Dataset (CSD):

Songs: "Erbarme dich" (ER), "Locus iste" (LI), "Nuper rosarum flores" (ND)
Singers: 4 different singers per voice part
Combinations: All possible singer combinations (4^4 = 256 per song)

Citation

If you use this dataset, please cite the original CSD paper and this preprocessing work.

License

CC-BY-4.0 - Same as the original Choral Singing Dataset