Deep Learning Models for Automatic Chord Recognition in Polyphonic Audio
Project Overview
This project investigated the long-tail problem in automatic chord recognition, where often existing systems default to predicting only major and minor triads when facing uncommon chords due to class imbalance in training data. A template-matching baseline was built and evaluated through a systematic ablation of preprocessing techniques. A Conformer-based architecture with structured multi-task heads for root, bass, and quality prediction was then proposed and implemented on the Bede supercomputer, representing the first application of the Conformer to this task. A real-time chord recognition iOS application, Amadeus, was also developed as a practical deployment target for the research.
What were the key results of your research project?
- Preprocessing has a significant impact on chord recognition accuracy. Tuning correction and harmonic/percussive source separation were the most impactful individual additions, and their combination yielded the strongest baseline performance.
- A novel Conformer-based architecture with structured multi-task heads was designed, implemented, and deployed on Bede.
- Amadeus, a real-time chord recognition iOS application, was built and is publicly available, demonstrating a practical application of such research.
GitHub Repositories
Conformer architecture for ARC implementation: https://github.com/cucuwritescode/conformer-acr
App at the research stage: https://github.com/cucuwritescode/amadeus
How do you feel you have benefitted from completing this internship and has it made you consider future career paths?
The internship gave me hands-on experience with HPC infrastructure, from writing SLURM job scripts to debugging distributed training across multiple GPUs. Working on Bede exposed me to the practical challenges of training deep learning models at scale, which is quite different from running things locally. It also strengthened my understanding of the full research pipeline, from literature review through to implementation and evaluation. The experience has reinforced my interest in pursuing research in machine learning and audio signal processing.