My research interests include automated analysis and synthesis of music compositions and performances: generation of style-specific accompaniment, interactive composition, and applications in music information retrieval.
In this project, I extend the automatic accompaniment system in my dissertation research by applying Decision Trees to chord tone determination, and by conducting two verification experiments. The automatic accompaniment system aims to assist music lovers in writing a complete song with sophisticated chord progressions. The system takes melody as input, and harmonizes it with style-specific accompaniments.
Phrase Structure Analysis in Expressive Performances
The objective of this research is to analyze the relations between the dynamics in expressive performances and musical phrase structure. We use an asymmetric quadratic curve to model the tempo changes in each phrase, and apply dynamic programming for seeking the best fit curves in order to define boundaries between phrases. The results of four piano performances are shown as follows.
This project develops a program that allows users to visualize the relations between tempo changes, loudness, and musical scores of real performances in audio recordings. Tempo tracking is provided by users through mouse clicks. The amplitude (loudness) of the audio, and the score obtained from MIDI, are shown. It can also be used as a tool for MIDI and audio alignment.
The audio key finding project designs a system which can automatically determine the most important music factor - key - from tonal music in polyphonic audio format. Various approaches have been experiments in this project to obtain pitches from audio recordings.I modified Professor Chew’s Spiral Array model, a mathematical model for tonality, adapting it to audio signals, which turned out to be the most effective method. I then proposed a fuzzy analysis method to clarify bass notes using information from higher frequency harmonics. My results showed that audio key finding can sometime outperform MIDI (discrete information) key finding, thus provided evidence to the research field that harmonics in music audio may be helpful for audio key finding.
In this project, we try to design a system to generate playable guitar scores from audio. This system contains two major parts. First we use signal-processing technique to identify the frequency/pitch of each input note. Then, based on the pitch and timing of each note, we apply an algorithm to generate guitar tabs which are the most playable ones for amateurs. We take audio in wave format as input, which is the most common music format. Experiments are conducted by having some pop guitar songs as examples to evaluate the performance of our system. Discussions are also provided on these experiments.
Projects I have conducted at USC:
•Automatic Style-Specific Accompaniment
•Phrase Structure Analysis in Expressive Performances
•Following a Pianist’s Paces
•Audio Onset Detection Using Machine Learning Techniques