| Summary: | "The linear prediction equations can be viewed as extremely simplified cases of the general Kalman filter theory. It would appear that if one were willing to pay a price in complexity, that some benefit should be received. Unfortunately, at the present in any case, the value of Kalman filter theory for the processing of real speech has not been demonstrated." (Markel & Gray, 1976). The aim of this thesis is to address concerns raised more than twenty years ago by Markel & Gray as in the above quotation. We place the Linear Prediction (LP) model of speech production in a Kalman filtering context. We show how it copes with the shortcomings of the more conventional LP methods in attempting to separate the glottal source and vocal tract filter. We also demonstrate how the Kalman filter estimate quality byproduct can be used to detect regions of the glottal closed phase. In an age where concerns regarding computational complexity are rapidly being eased, we believe that the time has come for more widespread use of the Kalman filter in speech processing. We have placed this research in the sphere of automatic speaker characterisation, but the potential of the Kalman filter extends far, far beyond.
|
|---|
