| Summary: | Speech signal processing is an old research topic within the communication theory community. The continously increasing telephony market brought special attention to the discipline during the 80’s and 90’s, specially in speech coding and speech enhancement, where the most significant contributions were made. More recently, due to the appearance of novel signal processing techniques, the standard methods are being questioned. Sparse representation of signals and compessed sensing made significant contributions to the discipline, through a better representation of signals and more efficient processing techniques. In this thesis, standard speech modeling techniques are revisited. Firstly, a representation of the speech signal through the line spectral frequencies (LSF) is presented, with a extended stability analysis. Moreover, a new Bayesian framework to time-varying linear prediction (TVLP) is shown, with the analysis of different methods. Finally, a theoretical basis for speech denoising is presented and analyzed. At the end of the thesis, the reader will have a broader view of the speech signal processing discipline with new insights that can improve the standard methodology.
|
|---|
