Robust Speaker Recognition Using Perceptual Stationary Wavelet Coefficients and Prosodic Feature in Noisy Conditions

Wavelet-based front-ends have been extensively utilized in speech processing systems, particularly for recognizing speech and speakers, and have significantly improved their performance. This work proposes a Robust Speaker Recognition (RSR) system based on a novel Perceptual Stationary Wavelet Coefficients PSWC and prosodic pitch feature for noisy environments. The PSWC approach combines a stationary wavelet packet, characterized by frequency band spacing similar to psychoacoustics Equivalent Rectangular Bandwidth scale, with the Implicit Weiner filtering (IWF). The use of IWF aims to minimize residual noise in each of 24 stationary wavelet sub-bands obtained adopted packet decomposition. The proposed RSR system’s performance utilizing PSWC features and its combination with pitch feature, which is estimated using Pitch Estimation Filter with Amplitude Compression technique, are evaluated under stationary and non-stationary noise conditions with heigh noise levels using the TIMIT dataset and Aurora noises. It is compared to that obtained using five traditional features Human Factor cepstral coefficients (HFCC), Inverse Mel Frequency Cepstral Coefficients (IMFCC), Bark Frequency Cepstral Coefficients (BFCC), Mel Frequency Cepstral Coefficients (MFCC), and Perceptual Linear Prediction (PLP) in terms of RSR accuracy and precision values and two additional measurements F1-score and recall. The obtained results using Gaussian Mixture Model-Universal Background Model demonstrate the superiority of our RSR system using the proposed combination of PSWC with pitch in various noisy cases. For instance, under Exhibition noise at 9 dB SNR, our RSR system achieved recognition accuracies of 58.73% using the proposed feature fusion (PSWC + pitch) and 47.54% using PSWC alone, while using IMFCC, BFCC, MFCC, PLP, and HFCC yielded accuracies of 32.06%, 13.73%, 16.83%, 17.86%, and 15.40%, respectively.