Classificação automática de categoria de cenas acústicas com redução de dimensionalidade baseada em projeções lineares e seleção de instâncias

Abstract

Acoustic scene classification consists of the task of recognizing the environment that an audio was recorded from its audio signal. The literature shows that deep learning-based methods provide excellent results. However, these methods have a high computational cost, making them unfeasible to use on limited devices. For this work we used the dataset provided by DCASE for the 2020 challenge Low-Complexity Acoustic Scene Classification. This challenge proposes a 500kb limit for storing model parameters. A baseline has also been published by the organizers of the challenge and consists of a neural network that achieved an accuracy of 87.30% with 450kb of parameters. To deal with this limit, two approaches were used that aim to reduce the size of the models for the task of classifying acoustic scene categories. The main goal of this work was to perform dimensionality reductions focused on cost optimization in the prediction phase. In the first approach evaluated, only the dimensionality reduction techniques PCA, RP, NMF were used with the KNN (K-Nearest Neighbors) and SVM (Support Vector Machine) classifiers. The best result that respected the 500KB limit was obtained with the NMF technique with 16 components and the SVM classifier. The F1-Score achieved was 86.64% with 390KB. In the second approach evaluated, it was also chosen to use the instance selection technique, since the classifiers used store the model from a sampling of the training set. In this approach, the dimensionality reduction techniques were combined together with the K-Means based instance selection technique (KMEANSC). The best result respecting the 500KB threshold was obtained with the PCA technique with 32 components using 15% of the training set, with the SVM classifier. The F1-score achieved in this situation was 87.84% with 410KB.