Róisín Loughran

Tonality driven piano compositions with Grammatical Evolution

We present a novel method of creating piano melodies with Grammatical Evolution (GE). The system employs a context free grammar in combination with a tonality-driven fitness function to create a population of piano melodies. The grammar is designed to create a variety of styles of musical events within each melody such as runs, arpeggios, turns and chords without any a priori musical information in regards to key or time signature. The fitness of the individuals is calculated as a measure of their tonality defined by a statistical distribution of the pitches in each piece. A number of short compositions are presented demonstrating that our system is capable of creating music that is interesting and unpredictable.

Musical instrument identification using Principal Component Analysis and Multi-Layered Perceptrons

This study aims to create an automatic musical instrument classifier by extracting audio features from real sample sounds. These features are reduced using Principal Component Analysis and the resultant data is used to train a Multi-Layered Perceptron. We found that the RMS temporal envelope and the evolution of the centroid gave the most interesting results of the features studied. These results were found to be competitive whether the scope of the data was across one octave or across the range of each instrument.

Genetic programming for musical sound analysis

This study uses Genetic Programming (GP) in developing a classifier to distinguish between five musical instruments. Using only simple arithmetic and boolean operators with 95 features as terminals, a program is developed that can classify 300 unseen samples with an accuracy of 94%. The experiment is then run again using only 14 of the most often chosen features. Limiting the features in this way raised the best classification to 94.3% and the average accuracy from 68.2% to 75.67%. This demonstrates that not only can GP be used to create a classifier but it can be used to determine the best features to choose for accurate musical instrument classification, giving an insight into timbre.

Speaker Verification on Unbalanced Data with Genetic Programming

Automatic Speaker Verification (ASV) is a highly unbalanced binary classification problem, in which any given speaker must be verified against everyone else. We apply Genetic programming (GP) to this problem with the aim of both prediction and inference. We examine the generalisation of evolved programs using a variety of fitness functions and data sampling techniques found in the literature. A significant difference between train and test performance, which can indicate overfitting, is found in the evolutionary runs of all to-be-verified speakers. Nevertheless, in all speakers, the best test performance attained is always superior than just merely predicting the majority class. We examine which features are used in good-generalising individuals. The findings can inform future applications of GP or other machine learning techniques to ASV about the suitability of feature-extraction techniques.

Clustering Agents for the Evolution of Autonomous Musical Fitness

This paper presents a cyclical system that generates autonomous fitness functions or Agents for evolving short melodies. A grammar is employed to create a corpus of melodies, each of which is composed of a number of segments. A population of Agents are evolved to give numerical judgements on the melodies based on the spacing of these segments. The fitness of an individual Agent is calculated in relation to its clustering of the melodies and how much this clustering correlates with the clustering of the entire Agent population. A preparatory run is used to evolve Agents using 30 melodies of known ‘clustering’. The full run uses these Agents as the initial population in evolving a new best Agent on a separate corpus of melodies of random distance measures. This evolved Agent is then used in combination with the original melody grammar to create a new melody which replaces one of those from the initial random corpus. This results in a complex adaptive system creating new melodies without any human input after initialisation. This paper describes the behaviour of each phase in the system and presents a number of melodies created by the system.

Feature selection for speaker verification using genetic programming

We present a study examining feature selection from high performing models evolved using genetic programming (GP) on the problem of automatic speaker verification (ASV). ASV is a highly unbalanced binary classification problem in which a given speaker must be verified against everyone else. We evolve classification models for 10 individual speakers using a variety of fitness functions and data sampling techniques and examine the generalisation of each model on a 1:9 unbalanced set. A significant difference between train and test performance is found which may indicate overfitting in the models. Using only the best generalising models, we examine two methods for selecting the most important features. We compare the performance of a number of tuned machine learning classifiers using the full 275 features and a reduced set of 20 features from both feature selection methods. Results show that using only the top 20 features found in high performing GP programs led to test classifications that are as good as, or better than, those obtained using all data in the majority of experiments undertaken. The classification accuracy between speakers varies considerably across all experiments showing that some speakers are easier to classify than others. This indicates that in such real-world classification problems, the content and quality of the original data has a very high influence on the quality of results obtainable.

Grammatical Music Composition with Dissimilarity Driven Hill Climbing

An algorithmic compositional system that uses hill climbing to create short melodies is presented. A context free grammar maps each section of the resultant individual to a musical segment resulting in a series of MIDI notes described by pitch and duration. The dissimilarity between each pair of segments is measured using a metric based on the pitch contour of the segments. Using a GUI, the user decides how many segments to include and how they are to be distanced from each other. The system performs a hill-climbing search using several mutation operators to create a population of segments the desired distances from each other. A number of melodies composed by the system are presented that demonstrate the algorithms ability to match the desired targets and the versatility created by the inclusion of the designed grammar.

Limitations from Assumptions in Generative Music Evaluation

The merit of a given piece of music is difficult to evaluate objectively; the merit of a computational system that creates such a piece of music may be even more so. In this article, we propose that there may be limitations resulting from assumptions made in the evaluation of autonomous compositional or creative systems. The article offers a review of computational creativity, evolutionary compositional methods and current methods of evaluating creativity. We propose that there are potential limitations in the discussion and evaluation of generative systems from two standpoints. First, many systems only consider evaluating the final artefact produced by the system whereas computational creativity is defined as a behaviour exhibited by a system. Second, artefacts tend to be evaluated according to recognised human standards. We propose that while this may be a natural assumption, this focus on human-like or human-based preferences could be limiting the potential and generality of future music generating or creative-AI systems.