핵심 개념
The author presents an unsupervised learning method for harmonic analysis based on a hidden semi-Markov model, focusing on chord quality templates and neural networks. The approach aims to automate harmonic analysis without the need for expensive labeled data or manual rule elaboration.
초록
The paper introduces a novel unsupervised learning method for harmonic analysis using a hidden semi-Markov model. It emphasizes chord quality templates and neural networks to automate the process efficiently. Despite some limitations in performance compared to supervised models, the proposed approach shows promise in advancing unsupervised harmonic analysis.
The study explores automated harmonic analyses, highlighting rule-based and supervised learning methods. Unsupervised harmonic analysis is presented as a challenging task due to simultaneous identification of keys and chords. Deep latent variable models are utilized to approximate probability distributions effectively.
Key metrics or important figures used:
Accuracy scores: JSBChorales60 - 66.8%, JSBChorales371 - 74.2%
Training epochs: Phase 1 - 480, Phase 2 - 240
Minibatch sizes: JSBChorales60 - 2, JSBChorales371 - 8
How can the proposed unsupervised learning method be enhanced to improve accuracy in harmonic analysis?
What are the implications of not considering enharmonic notes in the model's evaluation?
How might incorporating borrowed chords impact the model's ability to detect modulations accurately?