HARP: Harmonic-Aware Residual Partitioning for Neural Audio Codecs

HARP is a neural audio codec that partitions residual vector quantization across harmonically meaningful frequency bands to achieve high-quality, variable-bitrate audio compression. The key innovation is a harmonic-aware partitioning that guides the model to distribute codebook capacity across perceptually meaningful frequency bands, so a single model serves multiple bitrates by decoding a growing number of codebook groups.

This repository provides the official code for training and evaluating HARP.

Installation

Requires uv and FFmpeg (for audio decoding):

• uv: curl -LsSf https://astral.sh/uv/install.sh | sh
• FFmpeg: sudo apt install ffmpeg (Linux) or brew install ffmpeg (macOS)

git clone https://github.com/QiaoyuYang/harp-codec.git && cd harp-codec
uv sync                    # create .venv and install dependencies
source .venv/bin/activate

Pretrained models

Pretrained HARP weights are on the Hugging Face Hub. Download the checkpoint to the default location:

uv pip install "huggingface_hub[cli]"
hf download KelvinYang/harp-codec harp.ckpt --local-dir checkpoints

Usage

Inference on a single audio file

Reconstruct any audio file (wav/mp3/flac/...):

# full-rate reconstruction
python entry.py -i --input path/to/audio.wav --output recon.wav

# lower bitrate: fewer groups (1..4)
python entry.py -i --input path/to/audio.wav --n-groups 2 --output recon_2g.wav

Each run prints SI-SDR, multi-scale mel loss, LSD, and SNR against the input, and writes the reconstructed audio to --output.

Evaluate all bitrate tiers on one file

# metrics for every group tier
python entry.py -i --input audio.wav --eval-tiers

Evaluate on a dataset

Point the config's data.dataset_root (or --dataset) at a prepared dataset split:

# across all tiers, saving a JSON summary
python entry.py -i --eval-dataset --all-tiers \
  --dataset Jamendo --split val --output results_harp.json

Supported dataset types: Jamendo, LibriTTS, MUSDB18. Reported metrics are SI-SDR, multi-scale mel loss, LSD, and SNR.

Training

python entry.py -t --model harp --config harp/configs/train_harp.yaml

Before training, edit the config and set data.dataset_root to your audio directory and train.logdir to where checkpoints/logs should go. Training uses PyTorch Lightning; checkpoints and TensorBoard logs are written under logdir.

After training, export an inference checkpoint with scripts/export_weights.py. It keeps only the model weights, dropping the optimizer, discriminator, and callback state:

python scripts/export_weights.py path/to/harp-epoch=NN.ckpt checkpoints/harp.ckpt

Command-line arguments

Bitrate tiers

With 9 codebooks @ 1024 entries and a ~86 Hz frame rate:

Citation

@inproceedings{harp2026,
  title     = {HARP: Harmonic-Aware Residual Partitioning for Neural Audio Codecs},
  author    = {Yang, Qiaoyu and He, Lixing and Deng, Binyue and Zhao, Weifeng},
  booktitle = {Interspeech},
  year      = {2026}
}

Acknowledgments

This project builds upon the Descript Audio Codec (DAC) and uses the audiotools library for audio processing.

License

MIT