Hybrid methods for music analysis and synthesis: audio key finding and automatic style-specific accompaniment

UnrestrictedThis dissertation presents computational approaches to two problems in music analysis and synthesis -- audio key finding and automatic style-specific accompaniment. The two studies belong in the realm of computational music cognition, with a focus on automatic music analysis, which impacts information retrieval, and synthesis, which impacts the understanding of human creativity.; The first problem addresses the automatic extraction of key from audio signals. Audio key finding is challenging due to the unique characteristics of music audio, such as the harmonic series produced by individual tones. In this dissertation, I experiment with various approaches to obtain pitches from audio recordings, and discover that equal treatment of all detected pitches may not be the best approach. I then modify Chew's Spiral Array model, a mathematical model for tonality, to adapt it to audio signals. I also propose a fuzzy analysis method, inspired by the membership concept in fuzzy logic, to clarify bass notes using information from higher frequency harmonics. The results show that audio key finding can sometimes outperforms MIDI (symbolic) key finding, thus providing evidence that harmonics in music audio may be helpful for audio key finding. Sensitivity analysis on different system components reveal successful key finding strategies.; The second problem is concerned with automatic style-specific accompaniment (ASSA). The goal is to create a system that can analyze and learn the style from a small number of songs in the user’s preferred style, then automatically generate a sequence of chords to accompany the melody created by the user. The accompaniment generation process uses machine learning techniques to model melody-chord relations, and applies statistical modeling to chord transitions represented as neo-Riemannian transforms. I propose various evaluation methods, including the conducting of Turing tests. I also design several metrics to measure style specificity of generated accompaniments. I test different training strategies, and show that more data may not necessarily improve the style emulation results. System comparisons show that the accompaniments generated by the ASSA system are stylistically closer to the original than those generated by the Temperley-Sleator Harmonic Analyzer and a random chord generator