Polyphonic music generation using neural networks
In this project, the application of generative models for polyphonic music generation is investigated. Polyphonic music generation falls into the field of algorithmic composition, which is a field that aims to develop models to automate, partially or completely, the composition of musical pieces. Th...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis de maestría |
| Fecha de publicación: | 2021 |
| País: | España |
| Institución: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/361950 |
| Acceso en línea: | https://hdl.handle.net/2117/361950 |
| Access Level: | acceso abierto |
| Palabra clave: | Computer composition Neural networks (Computer science) Composició musical per ordinador Xarxes neuronals (Informàtica) Música -- Matemàtica Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial |
| Sumario: | In this project, the application of generative models for polyphonic music generation is investigated. Polyphonic music generation falls into the field of algorithmic composition, which is a field that aims to develop models to automate, partially or completely, the composition of musical pieces. This process has many challenges both in terms of how to achieve the generation of musical pieces that are enjoyable and also how to perform a robust evaluation of the model to guide improvements. An extensive survey of the development of the field and the state-of-the-art is carried out. From this, two distinct generative models were chosen to apply to the problem of polyphonic music generation. The models chosen were the Restricted Boltzmann Machine and the Generative Adversarial Network. In particular, for the GAN, two architectures were used, the Deep Convolutional GAN and the Wasserstein GAN with gradient penalty. To train these models, a dataset containing over 9000 samples of classical musical pieces was used. Using a piano-roll representation of the musical pieces, these were converted into binary 2D arrays in which the vertical dimensions related to the pitch while the horizontal dimension represented the time, and note events were represented by active units. The first 16 seconds of each piece was extracted and used for training the model after applying data cleansing and preprocessing. Using implementations of these models, samples of musical pieces were generated. Based on listening tests performed by participants, the Deep Convolutional GAN achieved the best scores, with its compositions being ranked on average 4.80 on a scale from 1-5 of how enjoyable the pieces were. To perform a more objective evaluation, different musical features that describe rhythmic and melodic characteristics were extracted from the generated pieces and compared against the training dataset. These features included the implementation of the Krumhansl-Schmuckler algorithm for musical key detection and the average information rate used as an estimator of long-term musical structure. Within each set of the generated musical samples, the pairwise cross-validation using the Euclidean distance between each feature was performed. This was also performed between each set of generated samples and the features extracted from the training data, resulting in two sets of distances, the intra-set and inter-set distances. Using kernel density estimation, the probability density functions of these are obtained. Finally, the Kullback-Liebler divergence between the intra-set and inter-set distance of each feature for each generative model was calculated. The lower divergence indicates that the distributions are more similar. On average, the Restricted Boltzmann Machine obtained the lowest Kullback-Liebler divergences. |
|---|