Automatic transcription of music using deep learning techniques

Gil, André Ferreira

http://hdl.handle.net/10400.8/4041

Use this identifier to reference this record.

Name:	Description:	Size:	Format:
André Ferreira Gil - Report_Model.pdf		2.7 MB	Adobe PDF	Download

Send Feedback

Authors

Gil, André Ferreira

Advisor(s)

Grilo, Carlos Fernando Almeida

Domingues, Patrício Rodrigues

Reis, Gustavo Miguel Jorge

Abstract(s)

Music transcription is the problem of detecting notes that are being played in a musical piece. This is a difficult task that only trained people are capable of doing. Due to its difficulty, there have been a high interest in automate it. However, automatic music transcription encompasses several fields of research such as, digital signal processing, machine learning, music theory and cognition, pitch perception and psychoacoustics. All of this, makes automatic music transcription an hard problem to solve. In this work we present a novel approach of automatically transcribing piano musical pieces using deep learning techniques. We take advantage of deep learning techniques to build several classifiers, each one responsible for detecting only one musical note. In theory, this division of work would enhance the ability of each classifier to transcribe. Apart from that, we also apply two additional stages, pre-processing and post-processing, to improve the efficiency of our system. The pre-processing stage aims at improving the quality of the input data before the classification/transcription stage, while the post-processing aims at fixing errors originated during the classification stage. In the initial steps, preliminary experiments have been performed to fine tune our model, in both three stages: pre-processing, classification and post-processing. The experimental setup, using those optimized techniques and parameters, is shown and a comparison is given with other two state-of-the-art works that apply the same dataset as well as the same deep learning technique but using a different approach. By different approach we mean that a single neural network is used to detect all the musical notes rather than one neural network per each note. Our approach was able to surpass in frame-based metrics these works, while reaching close results in onset-based metrics, demonstrating the feasability of our approach.