Optimized Deep Neural Networks Audio Tagging Framework for Virtual Business Assistant

Abstract

— A virtual assistant has a huge impact on business and an organizations development. It can be used to manage customer relations and deal with received queries, automatically reply to e-mails and phone calls.Audio signal processing has become increasingly popular since the development of virtual assistants. Deep learning and audio signal processing advancements have dramatically enhanced audio tagging. Audio Tagging (AT) is a challenge that requires eliciting descriptive labels from audio clips. This study proposes an Optimized Deep Neural Networks Audio Tagging Framework for Virtual Business Assistant to categorize and analyze audio tagging. Each input signal is used to extract the various audio tagging features. The extracted features are input into a neural network to carry out a multi-label classification for the predicted tags. Optimization techniques are used to improve the quality of the model fit for neural networks. To test the efficiency of the framework, four comparison experiments have been conducted between it and some of the others. From these results, it was concluded that this framework is better than the others in terms of efficiency. When the neural network was trained, Mel-Frequency Cepstral Coefficient (MFCC) features with Adamax achieved the best results with 93% accuracy and a 0.17% loss. When evaluating the performance of the model for seven labels, it achieved an average of precision 0.952, recall 0.952, F-score 0.951, accuracy 0.983, and an equal error rate of 0.015 in the evaluation set compared to the provided Detection and Classification of Acoustic Scenes and Events (DSCASE) baseline where he achieved and accuracy of 72.5% and