A speaker identification-verification approach for noise-corrupted and improved speech using fusion features and a convolutional neural network

The degraded quality of the speech input signal has a negative impact on speaker recognition techniques. We address the issues of speaker recognition from noise-corrupted audio signals in the presence of four noise variants, including factory noise, car noise, street traffic noise, and voice babble noise, as well as noise-suppressed enhanced speech. The goal of this research is to create a speaker recognition algorithm that is resistant to a diverse spectrum of speech capture quality, background scenarios, and interferences. In this work, three distinct features, including Mel Frequency Cepstral Coefficients (MFCC), Normalized Pitch Frequency (NPF), and Normalized Phase Cepstral Coefficients (NPCC) are combined. The analysis that MFCC, NPF, and NPCC illustrate distinct features of speech underlies our observation. A Convolutional Neural Network (CNN) is used in our speaker recognition strategy to learn speaker-dependent attributes from fragments of Mel features, normalized pitch features, and phase cepstral features of clean speech, corrupted speech, and enhanced speech. The performance is measured using the ITU-T test signals and compared to previous algorithms at different Signal-to-Noise-Ratios of 0 dB, 5 dB, 10 dB, and 15 dB. For enhanced speech, all three features, MFCC, NPF, and NPCC, provided productive speaker identification and verification performance.