🎶 From Discord to Harmony: Decomposed Consonance-Based Training for Improved Audio Chord Estimation

This repository contains the official implementation of the paper:

Poltronieri, A., Serra, X. and Rocamora, M. (2025) ‘From Discord to Harmony: Decomposed Consonance-Based Training for Improved Audio Chord Estimation’, in Proceedings of the 26th International Society for Music Information Retrieval Conference (ISMIR). Daejeon, South Korea: International Society for Music Information Retrieval, ISMIR 2025. Available at: https://arxiv.org/abs/2509.01588.

📦 Installation

1. Clone the repository:

   git clone https://github.com/andreamust/consonance-ACE.git
   cd consonance-ACE

2. Create the conda environment (Python 3.11):

   conda create -n ace python=3.11
   conda activate ace

3. Install dependencies:

   pip install -r requirements.txt

📂 Dataset Preparation

Our experiments use chord annotations and audio from:

• Training / Validation:

  • Isophonics dataset
  • McGill Billboard corpus

• Testing:

  • RWC Pop dataset
  • USPop dataset

Chord annotations are sourced from ChoCo: the Chord Corpus.

Prepare the dataset features by running:

python ACE/preprocess/preprocess_data.py

This generates cache files containing both the preprocessed audio features and chord labels. Parameters for the preprocessing script can be adjusted in ACE/preprocess/dataset.gin.

🚀 Training

To train a model, run:

python -m ACE.trainer --model model_name --name run_name

• --model: choose between

  • conformer: baseline classification model
  • conformer_decomposed: decomposition-based proposed in the paper.

• --name: optional, used for logging the run on Weights & Biases.

Parameters for the training script can be adjusted in ACE/trainer.gin.

🎯 Models

Two models are implemented in this repository:

• conformer: baseline architecture for chord classification (170 classes).
• conformer_decomposed: our proposed model, predicting root, bass, and pitch activations separately.

The conformer_decomposed model introduces several key innovations:

1. Decomposed Output Heads: Instead of a single output layer for all chord classes, we use separate heads for root, bass, and chord tones, allowing for more specialized learning.
2. Consonance Label Smoothing: adds music-theory informed smoothing, improving robustness and harmonic plausibility.

Models are stored in the ACE/models directory, which also contains .gin configuration files for each model.

🔮 Inference

Run chord inference on an audio file using the trained conformer_decomposed model:

python -m ACE.inference --audio path/to/audio.wav --out path/to/output.lab --chord-min-duration 0.5

The --chord-min-duration parameter sets the minimum duration (in seconds) for chord segments; shorter segments will be merged with adjacent ones.

This processes the entire track in 20 s segments, decodes predictions, merges identical consecutive chords, and outputs a .lab file:

start_time    end_time    chord_label
0.000000      2.581995    E:maj
2.581995      5.163990    B:maj
5.163990      7.745985    A:maj

The resulting .lab file contains the full chord sequence for the input track and can be visualised in Sonic Visualiser or any standard chord annotation tool.

By default, inference uses the pretrained checkpoint located at ACE/checkpoints/conformer_decomposed_smooth.ckpt, provided in this repository.

📑 Citation

If you use this code, please cite:

@inproceedings{poltronieri2025discord,
  title     = {From Discord to Harmony: Decomposed Consonance-Based Training for Improved Audio Chord Estimation},
  author    = {Poltronieri, Andrea and Serra, Xavier and Rocamora, Martín},
  booktitle = {Proceedings of the 26th International Society for Music Information Retrieval Conference (ISMIR)},
  publisher = {International Society for Music Information Retrieval, ISMIR 2025},
  year      = {2025},
  location  = {Daejeon, South Korea},
  url={https://arxiv.org/abs/2509.01588}, 
}

📜 License

MIT License

Copyright (c) 2025 Andrea Poltronieri, Xavier Serra, Martín Rocamora

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.