Deep neural networks-based systems for acoustic scene recoginition

Abstract

Acoustic scene classification (ASC) refers to the identification of the environment in which audio excerpts have been recorded, which associates a semantic label to each audio recording. This task has drawn lots of attention during the past several years as a result of machines and electronics such as smartphones, autonomous robots, or security systems acquiring the ability to perceive sounds. This work aims to classify 10 common indoor and outdoor locations using environmental sounds. To accomplish the ensuing task, we have performed multiple experiments using a dataset which consists of 14400 sound files. The goal is to explore three different aspects of an ASC system: deep learning architecture, feature extraction technique and data augmentation method. In particular, two deep neural networks have been employed in the construction of our systems namely: Residual Neural Network (ResNet) and Alex Neural Network (AlexNet), along with a combination of feature representations based on signal processing techniques. Specifically, 3 feature sets have been extracted: Log-Mel energies, ∆Log-Mel energies and ∆∆Log-Mel energies. Moreover, this work deeply explores the use of Mixup data augmentation method and the effects of varying its hyperparameters in reducing the chance of overfitting. A series of thorough experimental comparisons and statistical tests have been performed with regards to evaluating our systems. The obtained results indicate that a proper choice of the feature set is crucial in view of the deep learning architecture. Additionally, statistical testing has shown the significant impact of mixup data augmentation technique on the predictive performance of our models, as systems trained on augmented data have a considerably better generalization ability compared to the counterpart systems trained on original data. Moreover, we have found that a well-tuned mixup hyperparameter α significantly improves the classification system performance. Keywords: Acoustic Scene Classification, Feature Extraction, Data Augmentation, Deep Learning, Mixup, Statistical Tests.