Automatic guitar performance assessment: datasets, algorithms and metrics

Abstract

The aim of this project is to contribute to the development of a music transcription system of guitar performances in the context of music education. Using the Music Critic system as a baseline we identified aspects to improve and then we implemented and tested our contributions. We show how the system reacts to different contextual variations and discuss possible repercussions on a real context application. Also, we studied how the model used behaves using different training conditions. Furthermore, the Five Guitar dataset, with 90 guitar recordings, is designed especially for this project and publicly available. We use a data augmentation strategy to obtain a higher number of recordings simulating different rooms, mics, effects and recording setups. We observe that room acoustics and recording setup could generate biases on the final performance of the model and that the system is consistent according to timber. Also, we discover the need of representing all the pitch-class sets into the training set, which could be a limitation in a real situation, plus a high bias in the model.