Technical reports

Simulating music learning with autonomous listening agents: entropy, ambiguity and context

Ben Y. Reis

September 1999, 200 pages

This technical report is based on a dissertation submitted July 1999 by the author for the degree of Doctor of Philosophy to the University of Cambridge, Queens’ College.

Abstract

Music learning describes the gradual process of acculturation through which listeners in different cultures develop diverse sets of musical preferences and intuitions. This dissertation describes Maestro, a system designed over the course of this research to simulate certain aspects of music listening and learning.

In order to maintain the unbaised flexibility necessary for handling music from different styles, Maestro does not incorporate any a priori style-specific knowledge into its design. Instead, Maestro is based on a bottom up approach that maximises the use of perceptual information present in a performance.

Maestro’s operation involves four stages: it first segments a musical performance on-line according to perceptual cues (segmentation) and constructs an appropriate model of the performance (modelling), based on the context modelling paradign. This model is simultaneously used to generate expectations about upcoming events (prediction) and to interpret events once they have arrived (parsing).

Ambiguity is an essential part of music listening, especially in the context of learning, and can cause multiple hypotheses of interpretation to arise. A novel multi-agent methodology is developed and incorporated into Maestro for generating, maintaining, and reconciling these hypotheses. An information theoretic approach, based on measuring two types of entropy, is used to objectively evaluate the system’s relative prediction performance. It is also found that entropy, along with a measure of agent activation, is useful for identifying and classifying different types of ambiguity.

Experiments performed with a collection of 100 Bach chorale melodies provides a basis for comparison with previous machine modelling reseasrch and with data from human subjects. A much larger collection of roughly 8,000 folk songs from different cultures enables significant large scale and panstylistic music learning experiments to be performed. Perceptually guided segmentation is argued to yield more cognitively realistic context models than other methods, and it is also empirically shown to yield more efficient models for prediction. Additionally, an adaptive modelling strategy allows appropriate multiple-step-ahead predictions to be generated. Finally a distributed, agent-based parsing methodology is developed and implemented.

The system provides insights into what implications certain theories from cognitive musicology have when put into practice. Maestro’s flexible design together with the range of experiments performed and the diverse corpus of musical data enable a thorough and systematic machine-simulated study of key aspects of musical learning to be carried out.

Full text

PDF (16.8 MB)

BibTeX record

@TechReport{UCAM-CL-TR-472,
  author =	 {Reis, Ben Y.},
  title = 	 {{Simulating music learning with autonomous listening
         	   agents: entropy, ambiguity and context}},
  year = 	 1999,
  month = 	 sep,
  url = 	 {https://www.cl.cam.ac.uk/techreports/UCAM-CL-TR-472.pdf},
  institution =  {University of Cambridge, Computer Laboratory},
  doi = 	 {10.48456/tr-472},
  number = 	 {UCAM-CL-TR-472}
}