Modelling the Flow Inherent in Speech Representations

This paper presents two new methods for modelling the flow inherent in speech: flow-based prediction (FBP)and acoustic flow interpolation (AFI). These are presented as extensions of the form of prediction implied in calculating the delta and delta-delta coefficients often used in automatic speech recognition.All these methods are presented as special cases of a general vector linear prediction model, but it is shown that the new techniques, which make the flow of features within the data explicit, are significantly better at modelling spectrogram-like data. Several speech representations, using both parametric and non-parametric analyses, are discussed both in terms of their ability to represent speech accurately and of their appropriateness to these flow-based models. AFI and FBP error coefficients, for both male and female speakers, are measured and compared with the delta and delta-delta coefficients. Wherever possible, the parameters and methods used to produce the representations have been chosen to be directly comparable with one another.