Speech enhancement system using deep neural network optimized with Battle Royale Optimization

• Published in: Biomedical signal processing and control 2024-06, Vol.92, p.105991, Article 105991
• Main Authors: Kumar Shukla, Neeraj, Shajin, Francis H, Rajendran, Radhika
• Format: Article
• Language: English
• Subjects: Battle Royale Optimization (BRO) algorithm; Deep neural network (DNN); Kalman filter (KF); MATLAB; Speech enhancement
• ISSN: 1746-8094; 1746-8108
• DOI: 10.1016/j.bspc.2024.105991

•A DNN based Speech Enhancement system using BRO is proposed.•To enhance the quality and increases the SNR for speech signal.•The intention of the proposed work is to improve the speech quality.•The input signals are taken from TIMIT dataset.•Objective is to enhance the robustness to the unseen noise types. Speech enhancement (SE) is widely employed in various fields, like speech recognition, wireless communication, hearing aids, smart home devices. In this manuscript, a Speech Enhancement System using Deep Neural Network Optimized with Battle Royale Optimization is proposed in this manuscript to enhance the quality and increases the SNR for speech signal. The intention of the proposed work is to improve the speech quality and intelligibility using DNN and Battle Royale Optimization. Initially, the input signals are taken from TIMIT dataset. DNN-based SE is used to enhance the speech quality and mostly operates in the frequency domain. DNN based Kalman filter (KF) is proposed in this work for time domain SE. The objective of this manuscript is “to enhance the robustness for unseen noise types in speech signal and increase the speech quality”. The speech enhancement based on DNN with Battle Royale Optimization(BRO) method is proposed to optimize the objective function. The proposed approach can function reliably in both matched and mismatched acoustic settings by utilizing both standard KF and DNN-based reconstruction. The simulation procedure is implemented in MATLAB platform. The simulation outcomes depicts that the proposed approach achieves 16.04%, 21.83%, and 24.32% higher Short-Time Objective Intelligibility (STOI), 25.38%, 31.19%, and 20.24% greater PESQ and 29.07%, 24.90% and 34.22% lower MSE compared with existing models, like Regularized sparse features for noisy SE utilizing deep neural networks (PADNN-SES), Multi-channel SE depending on early and late fusion convolutional neural networks (CNN-SES) and Self-attending RNN for speech enhancement to improve cross-corpus generalization (SARNN-SES) respectively.