ABSTRACT
Single-channel speech enhancement relies upon differences between the characteristics of speech and noise. For some speech processing, such as pitch detection, longer frames are used to encompass multiple pitch periods. This chapter provides a discussion on speech enhancement and looks at how an autoregressive model for a signal may be formulated in terms of speech frames. For speech, the statistics can be considered stationary within each frame. However, they will vary from frame to frame. To obtain a better representation of the statistics of speech, one must also consider the relationships between frames. The chapter considers some specific unitary transforms used in speech enhancement. If the speech has been generated by an autoregressive process the autocorrelation function and hence the power spectrum have a limited number of free parameters. The iterative Weiner filter is an example of an enhancement algorithm that uses a parametric model for the power spectrum of the speech.
