Classificação automática de cenas acústicas usando algoritmos de clusterização

Abstract

The Acoustic Scene Classification problem deals with assigning an environment-related label to an audio signal. Among the labels are parks, airports, streets and public squares. In this work we present four approaches to this problem based on machine learning and digital signal processing. Our main objective was to minimize the computing power required for model training and making predictions, while keeping classification performance at acceptable levels. Our highest performing method consists in describing audios with Mel-Frequency Cepstral Coefficients and then grouping them with a 2-level K-means clustering approach. This clustering approach describes classes using sounds that are common among audios of the same class. This promotes generalization and lowers the number of data points needed for model training. In turn, this lowers the system computing power requirements. This approach reduced the number of data points to around 10% of the total, and achieved 62% accuracy in the DCASE 2018 Task 1a dataset. This result is comparable with the results obtained by the baseline system, which is based on convolutional neural networks.