Clasificador de escenas acústicas basado en redes neuronales artificiales y análisis de componentes principales

Abstract

Acoustic scene classification has been gaining importance in recent years. The applications are interesting and additionally, it represents a challenge to implement a computational tool that can detect complex and diverse sounds, such as those presented in real environments. In this work, convolutional neural networks and feed-forward are implemented, trained with individual characteristics such as Mel Frequency Cepstral Coefficients (MFCC), gamma tones and Discrete Fourier Transform (DFT), extracted to sounds in 100 ms windows with 50% overlap, then form segments of 1 and 10 seconds. Neural networks are also trained with combinations of characteristics (DFT-Gamma, DFT-MFCC, Gamma-MFCC, DFT-Gamma-MFCC). Subsequently, the number of input coefficients is reduced by applying PCA, verifying the impact of this reduction on the performance and training time of different neural network architectures. In both cases cross validation is used with 80% of the data for training and 20% for validation. This work was development using the DCASE2018 database.