Incorporating AM-FM effect in voiced speech for probabilistic acoustic tube model

A complete speech model can improve performance for many speech applications. Probabilistic Acoustic Tube (PAT) is a probabilistic generative model of speech that has been shown potentially useful in a number of speech processing tasks. A point missing in previous PAT models is that they overlook AM/FM effect in voiced speech, which is in fact common and non-negligible. In this paper, we significantly improve the voiced modeling of PAT with a probabilistic model of AM/FM effect, which is developed from Bayesian Spectrum Estimation method. Experiments show that the new PAT is able to fit the voiced speech spectrum with greater accuracy in the presence of AM/FM effect.