MulTTiPop Audio Pipeline

Download the gclef-cmu/multtipop snapshot and its referenced YouTube segments, apply FlashSR, and save 48 kHz audio to:

$DATA_ROOT/multtipop/audio/{ID}.{opus|mp3|flac}

The pipeline downloads only the time range specified in the metadata. It is resumable, validates outputs before publishing them, keeps dataset files as real files rather than symbolic links, and retries each download three times after the initial attempt.

Pipeline

Requirements

• Linux, an NVIDIA GPU, and a CUDA-compatible driver
• Conda, Git, npm, and Node.js
• FFmpeg/FFprobe with Opus, MP3, and FLAC encoders
• About 3.4 GB for FlashSR checkpoints, plus dataset and audio storage

Setup

git clone https://github.com/mimbres/multtipop-audio.git
cd multtipop-audio
bash setup_env.sh

The setup script creates a project-local .venv, installs a CUDA 12.6 PyTorch build and the pinned FlashSR revision, and installs a local Node.js 22 runtime. Set CONDA_BIN or ENV_PREFIX to override the detected Conda executable or environment location.

Run

export DATA_ROOT=/path/to/data
./run_pipeline.sh --data-root "$DATA_ROOT"

Opus is the default (libopus, 256 kbit/s VBR). MP3 uses 320 kbit/s CBR and FLAC uses lossless 24-bit encoding:

./run_pipeline.sh --data-root "$DATA_ROOT" --format mp3
./run_pipeline.sh --data-root "$DATA_ROOT" --format flac

Useful options:

--retries N             Retries after the initial download attempt (default: 3)
--cookies FILE          Netscape-format YouTube cookie file
--format FORMAT         opus, mp3, or flac
--id ID                 Process one ID; repeat to select multiple IDs
--limit N               Process the first N selected records
--keep-source           Keep downloaded source segments
--overwrite             Regenerate valid outputs
--skip-dataset-download Use an existing dataset snapshot
--device DEVICE         FlashSR device (default: cuda:0)

Rerun the same command after interruption; valid outputs are skipped. Progress and failure details are written under $DATA_ROOT/multtipop/logs/:

cat "$DATA_ROOT/multtipop/logs/status.json"
tail -f "$DATA_ROOT/multtipop/logs/pipeline.log"
cat "$DATA_ROOT/multtipop/logs/failed_ids.log"

If YouTube requires sign-in, update yt-dlp and pass exported Netscape-format cookies with --cookies. Never commit cookie files.

Reference run

A completed 572-record Opus run produced 557 valid files: 555 were processed and 2 already-valid files were skipped. The remaining 15 records failed during download after retries. Their IDs and failure stage are tracked in failed_ids.log. YouTube availability changes over time, so future results may differ.

Tests

python -m pip install -r requirements-dev.txt
pytest -q

Disclaimer

This is an unofficial repository and is not endorsed by the MulTTiPop authors or YouTube. All music and recording rights remain with their respective owners. Use this software only for research, comply with applicable licenses and laws, and proceed at your own risk.

Citation

Please cite the MulTTiPop paper:

@article{pruyne2026multtipop,
  title={MulTTiPop: A Multitrack Transcription Dataset for Pop Music},
  author={Pruyne, Nathan and Stoler, Benjamin and Chen, William and Huang,
          Chien-yu and Watanabe, Shinji and Donahue, Chris},
  journal={arXiv preprint arXiv:2607.08756},
  year={2026}
}