Noise resilient speech signal analysis using non-stationary signal processing techniques

Abstract

Speech signal processing applications have gradually found place in diverse elds such as mobile phones, text reader applications, GPS, human-computer interactions, wireless communications, voice pathology detection. Real time language translation and natural language interpretation are the new emerging areas that employ speech signal processing. The scope of speech signal processing applications is expected to expand and grow in the coming years.This thesis focuses on the noise resilient analysis of the speech signal using nonstationary signal processing techniques. Speech signal analysis in the low frequency range (LFR) is shown to be advantageous for robust determination of glottal characteristics pertaining to the voiced regions of a speech signal. It is useful for many applications such as text to speech synthesis, speaker recognition and emotion recognition. This thesis proposes noise resilient and accurate algorithms for instantaneous V/NV detection, extraction of the time-varying F0 component of a voiced speech signal and GCI identi cation. A novel technique for decomposition of a multi-component non-stationary signal (such as speech signal) into AM-FM mono-component signals is proposed in the last chapter of this thesis. It is employed for formant analysis of the voiced speech signal.The proposed V/NV detection algorithm exploits the property that in the LFR, the energy over the time-frequency plane is present only during voiced regions of the speech signal. The proposed iterative algorithm caters to the challenging problem of reliable extraction of the time-varying F0 component of a voiced speech signal in the presence of noise, without the need of time-varying lters. The proposed GCI identi cation method locates GCIs reliably and accurately by employing negative cycles of the extracted timevarying F0 component of a voiced speech signal to provide coarse estimate of the intervalswhere GCIs are likely to occur. Finally, a novel iterative decomposition approach is proposed to extract either only strong or strong cum weak AM-FM mono-component signals from a multi-component non-stationary signal (such as a voiced/unvoiced speech signal). The proposed iterative decomposition approach e ciently extracts the formant components of a voiced speech signal. The proposed iterative decomposition approach when used along with discrete energy separation algorithm (DESA) performs e cient and noise resilient formant analysis.