Detection of Vowels in Speech Signal with Application to Children s Keyword Spotting

Abstract

The vowel onset points (VOPs) and vowel end points (VEPs) are newlinetwo major events in a speech signal. The vowels can be classified newlinefrom other non-vowels anchoring VOPs and VEPs. In most cases, newlinethe signal transition at the VEP is slow compared to the VOP, which newlinealso depends on bounding sound units. Consequently, the detection of newlineVOPs and VEPs using a single algorithm is challenging. The detection newlineof these events is also affected by ambient noise. This thesis work aims newlineat extracting scalar features from the speech signal for simultaneously newlineenhancing the detection of VOP and VEP events for the classification newlineof vowels in a running speech. Then, by classifying each analysis frame newlineas vowel or non-vowel, a non-uniform spectral smoothing approach newlineis proposed for the development of a pitch-robust keyword spotting newlinesystem.The salient contributions of this thesis work are as follows: newlinei.) A novel temporal processing approach for extracting temporal newlinedynamics from speech signal for enhancing detection accuracy of newlineVOPs and VEPs through a single algorithm. newlineii.) An approach using variational mode decomposition (VMD) for newlinesuppressing the effect of ambient noise and reconstruction of the newlinesmoothed magnitude spectra to represent the resonance structure newlineof the vocal-tract system to detect VOPs and VEPs. newlineiii.) A robust scalar feature by extracting variation of short-term newlinemagnitude spectra (ST-MS). newlineiv.) A spectral processing approach for detection of vowels following newlinethe cepstral feature extraction process, to reduce the overall computation newlinein the development of speech-based applications using newlinevowel event. newlinev.) An approach for enhancing detection of VEPs using non-linear newlinemapping. newlinevi.) A non-uniform spectral smoothing approach using knowledge of newlinevowels for development of pitch robust KWS system.