SHO based Deep Residual network and hierarchical speech features for speech enhancement

The human frequently finds difficulty in understanding the speech due to the real-world noises. The presence of external noises corrupts the listening comfort of user. Hence there is a need for the enhancement of speech. In this paper, the Shepherd Harmony Optimization (SHO)-based Deep Residual network (DRN) is developed for speech enhancement. Here, the developed SHO-based DRN is the combination of the Shuffled Shepherd Optimization Algorithm (SSOA) and Harmony Search optimization (HS). The Hanning window is used for the pre-processing of the input data. In this method, the Bark Frequency Cepstral Coefficients (BFCC) and Fractional Delta amplitude modulation spectrogram (FD-AMS) are used for the feature extraction. Moreover, the noises present in speech signals are predicted for eliminating the distorted noises and external calamities. Besides, the DRN classifier is utilized to improve the speech signal. The classifier is trained by newly devised optimization algorithm. Besides, the developed speech enhancement technique obtained better performance in terms of Perceptual Evaluation of Speech Quality (PESQ) with 2.646 and Root Mean Square Error (RMSE) with 0.0067. © 2022, The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature.