AN ENHANCED AUDIO EVENT DETECTION WITH ATTENTION NEURAL NETWORKS

Abstract

Multimedia recordings are very vital in the aspect of audio event detection with attention neural networks and it’s a task of recognizing an audio event in an audio recording. The aim of the proposed work is to improve the development of existing audio/sound event detection in continuous streams and audio recording. Also, to compute the classes of audio events such as gunshots, screaming, door slamming, bell ringing, coffee, bird singing etc from an audio recording and also to estimate the onset and offset of these acoustic events. In this work, the propose system is going to use the modern machine learning methods called attention neural networks. The enhancement in the quality of audio event detection is achieved using an attention neural network based approach. Different activation functions that include RELU, LeakyReLU, ReLU6, ELU, and Swish known as SILU were investigated, the performance of the models using the above mentioned activation functions and Evaluate the performance of the baseline system using the different activation functions and compare the performance with the results of the existing studied papers were presented and discussed. As discussed and shown in this research, Swish network achieved mAP of 0.361432 and dprime of 2.642 outperformed the ReLU network and D-prime, from the baseline paper even though they both achieved the same AUC using the same architecture with 1024 hidden units. Using the feature level attention model, Swish activation function with mAP of 0.370 outperformed ReLU with mAP 0.369 in the baseline paper. Swish performance is...