Approaches for time-frequency analysis of speech signals without windowing
B. Yegnanarayana, International Institute of Information Technology (IIIT) Hyderabad
A new approach called ´zero-time liftering´ (ZTL) is proposed for analysis of speech signals. The term is used analogous to ´zero-frequency filtering´ (ZFF) method proposed earlier for extracting the excitation source information from speech signals. The nearly polynomial-type growth/decay of the spectrum masks the spectral features, if the ZTL is applied several times. But using the high resolution property of the numerator of group-delay function, the spectral features in the analysis segment can be extracted. The spectral information can be obtained at every instant of time, thus extracting the instantaneous spectral features. The method is illustrated for analysis of different categories of speech sounds. Some illustrations of analysis of speech sounds that require good temporal and spectral resolution will also be presented.
In this talk I will discuss the motivation for using the ZFF and ZTL methods for speech analysis, and show how these two methods complement each other in extracting the information about the excitation source and vocal tract system characteristics from speech signals, respectively.
Application of these methods for analysis of paralinguistic sounds like shout and laughter will also be discussed.