Piano Note Recognition: Classification Aided by Convolutional Neural Networks

Abstract

[ES] Una tarea que humanos pueden realizar con facilidad es el reconocimiento de notas de piano. Por ello he decidido usar una Red Neuronal Convolucional para enfrentar este problema. El entrenamiento se hizo vía un dataset que fue generado por software. Este problema también se expandió a casos de sonido de fondo aleatorio.

Dado que el dataset es muy robusto y es relativamente simple, el objetivo fue alcanzado con creces. En diferentes pruebas con distintos datasets, siempre la Red consigue predecir con una precisión mejor al 95%.

[EN] Artificial Neural Networks has changed how we solve problems that seemed unsolvable. One of such problems is audio recognition. The aim of the thesis was to recognize piano musical notes using neural networks. In this case, it was chosen the task of Note Identification, a task that most humans are not capable. It was tackled using a Convolutional Neural Network, as a problem of Supervised Learning. The Dataset was generated specially for this work, and in some cases, it was mixed with background rain noise. Even was expanded to the recognition of chords. Neural network efficiency was quantified as the recognition accuracy in unseen data. A Convolutional Neural Network was created that recognized piano notes. Accuracy was 100% for one note, and 96.94% for 3-note chords with background noise. In conclusion, Convolutional Neural Network allows the recognition of notes in environments with background noise. This approach could be used for more complex recognitions, including other instruments or sounds.