Chroma determination for musical chord identification

• Main Author: Ng, Nathanael Song Jin
• Other Authors: Saman S Abeysekera
• Format: Final Year Project
• Language: English
• Published: 2015
• Subjects: DRNTU::Engineering::Electrical and electronic engineering::Electronic systems::Signal processing
• Online Access: http://hdl.handle.net/10356/64331
• Institution: Nanyang Technological University

Automated chord identification provides a way of quickly and accurately identifying unknown chords. In this study, a simple and fast algorithm that can be scalable will be created for the purpose of chord identification. Analysis of the polyphonic audio will be conducted in the frequency domain. Past studies showed that the extraction of features from the frequency data can be done by utilising 12 chroma bins, called Pitch Class Profiles (PCP). Identification is done through matching these PCPs with Chord Type Templates (CTTs). However, the presence of noise often resulted in misidentified chords. Another study introduced the Harmonic Product Spectrum (HPS) and improved the PCPs by decimating the original magnitude spectrum to eliminate non-musical noise. In this study, different methods of chord identification and enhancing the results will be tested. Three approaches to weighting the CTTs, binary templates, manually calibrated templates or a hybrid of both methods, will be studied in order to determine the best approach. In addition, the optimum number of harmonics to decimate the magnitude spectrum by will be determined. Finally, the algorithm will be tested with the piano and guitar to determine its effectiveness across instruments. The results obtained showed that decimating the magnitude spectrum up till the second power of two was found to be the most effective at eliminating noise. When used in conjunction with the HPS, it was found that using the original binary templates and hybrid approach produced an accuracy of around 95%, but the hybrid approach produced an overall greater separation of scores. Furthermore, manual calibration was found to be unsuitable and not scalable. Finally, the overall accuracy of the algorithm when given real-world chords resulted in a minimum identification accuracy of 75% due to the effects of ambient noise. For future studies, it is recommended that more diverse chord types, such as extended chords, produced by a diverse range of instruments, be examined to facilitate the further refinement of the hybrid approach.