Using Machine Learning To Produce Expressive Musical Performance.

摘要

The use of artificial intelligence is common in the research of musicology, which involves the large-scale analysis of empirical data. Recent research studies show that it is possible to represent musical style in terms of local and global parameters. The local parameters arise from the performance of similar motifs in the phrase. The global parameters arise from the performance of the phrase as a whole. Performers tend to perform similar structures in a similar way. Based on these observations, we propose a method for reproducing the style parameters from music recordings. The pitch and beat were first extracted using a modified algorithm based on Peeter's [19] and Dixon's [21] algorithms, respectively. We then tracked the key by Krumhansl-Schmuckler's [22] algorithm. To predict the chord progression, we used a Hidden Markov Model (HMM) and chord transition matrix. To identify the phrases, we segmented the music by cadence, recurring pitch patterns, and local energy content. The phrases were then trained and re-targeted with a Support Vector Machine (SVM). The end result is a re-targeting of style parameters including dynamics, tempo and articulation. Experiments show that our method reproduces a performer's style with a high level of correlation to real performances. *Please refer to dissertation for footnotes.