Master of Science in Computer Science Thesis Defense

Location:
In Person - Tepper Building, Room 1403

Speaker:
HUIRAN YU, Masters Student, Computer Science Department, Carnegie Mellon University
https://www.linkedin.com/in/huiranyu

A Study of Statistical and Music-Theoretical Melody Prediction

Melody prediction is an essential research focus in computer music, aiming to predict melody terms given musical context. Melody prediction can help people understand how humans form melodic anticipation while listening and also contributes to the melody generation task in automatic composition. Nowadays, most studies only focus on developing new methods to model musical sequences. However, constructing effective techniques to measure model behavior also demands attention. In our research, we offer an information entropy metric that can be applied to standard models, then further combine music theory with models to see if we can get better outcomes.

We first established a metric to measure the capability of baseline models. Each model generates a probability distribution over terms in the sequence, and we calculate the average entropy throughout the melody.  Stronger models are likely to generate lower entropy, which means music is more predictable under these models. We found models trained on the whole dataset and those trained within the particular song show drastic differences. Surprisingly, training on a large dataset results in lower performance. After setting up the baseline, we designed another model recognizing periodic occurrences of notes and patterns, incorporating music characteristics of fixed phrase length and periodic repetition.

This simple model makes satisfying predictions, and with an ensemble strategy considering the entropy value and confidence of each model, we combined the new model with a statistic model reducing the prediction error from 7.5% to 6.5%, eliminating 13% of the failure cases.

Thesis Committee: Roger Dannenberg (Chair), Danny Sleator

Additional Information