Automatic music transcription using neural networks

Abstract

The use of artificial intelligence to solve problems that were not previously viable is growing exponentially. One of these problems is obtaining the musical notes (the music score) given a song in audio format. This task has a high complexity due to the large number of notes that can be played at the same time by different instruments. This project makes use of the Musicnet dataset which provides the audio data of 330 songs with their corresponding note labels. To extract relevant information and derive the features, Constant-Q Transform has been applied to transform the audio data to the frequency domain in a logarithmic scale. In addition, one-hot encoding vectors have been used to represent the output data, i.e., the music notes. Then, a deep neural network is trained to recognise the score given the music audio information. A research has been carried out to find the most appropriate methods to solve the problem. Besides, different topologies of neural networks have been developed to find which of them offers the best outcomes. The results obtained are positive since a high percentage of prediction accuracy has been achieved taking into account the great number of combinations that the problem presents.