Phase-based Technique for Signal Reconstruction and Its Applications in Speech Processing

Abstract

The problem of retrieving the phase of the signal from the magnitude of its Fourier transform is called as phase retrieval. There are several methods for signal reconstruction that have their own merits and demerits. In this thesis, we exploit the idea of encoding of the phase in the magnitude spectrum of the signal using cepstral processing. Using the concept of phase encoding, we propose a new algorithm for signal reconstruction. In particular, we consider the odd-even decomposition. We propose the shifting of the real part of the DTFT of signal (shifting in the frequency-domain) and make it non-negative. Using the proposed method, we try to improve the quality of the reconstructed signal. In addition, another significant contribution of this thesis is, we propose novel spectral features called as phase-encoded Mel cepstral coefficients (PEMCC). These features are used for spoofed speech detection (SSD) as well as automatic speaker verification (ASV) using the I-vector-based system. When the proposed feature set is fused with state-of-the-art MFCC (Mel frequency cepstral coefficients) feature set, it significantly reduces the equal error rate (EER) compared to the MFCC alone. This indicates the proposed feature captures the complementary information than the MFCC alone for both the applications consider in this thesis work.