Layer modified residual Unet++ for speech enhancement using Aquila Black widow optimizer algorithm

Speech enhancement techniques face computational demands, well-developed datasets, and better quality speech signals. Deep learners help deal with different noise types; still, the challenges offered by environmental noises require highly efficient and robust systems. This paper presents a lightweight deep-learning design with a heuristic-inspired model for generating an enhanced speech signal from noisy speech data. The model aims to remove different environmental noises affecting the speech signal. The noisy speech data are converted into spectrograms with Short-Time Fourier Transform (STFT). The noisy spectrogram is processed through the newly developed speech enhancement model namely, Layer Modified Residual Unet++ (LMResUnet++). The developed LMResUnet++ is designed through an atrous convolution layer, and it can capture multi-scale information without additional training parameter requirements. Also, the design is made compactable through the proposed hybrid optimization algorithm namely, Aquila Black Widow Optimization (ABWO), and it optimizes various hyperparameters of the developed LMResUnet++. The final denoised spectrogram from the LMResUnet++ undergoes Inverse STFT, and the final enhanced speech signal is restored. Further, different experiments are held to prove the efficacy of the system. Results prove that the developed LMResUnet++ achieved PESQ values of 7.93%, 5.75%, 3.86%, and 1.90% improved than DeepUnet, MTCNN, STCNN, and ResUnet++ respectively.