Speech production based on lossy tube models: unit concatenation and sound transitions

The discrete time tube model is well established in speech analysis and synthesis providing a simplified modeling of the vocal tract. The standard lossless tube model is extended by introducing frequency dependent losses. In this contribution it is shown how the lossy vocal tract model can be used for speech production. For that purpose the vocal tract areas of this model can be estimated from speech signals by an optimization algorithm. The estimated model parameters can be used successfully for resynthesis. Furthermore speech units are concatenated by transitions of the vocal tract areas. For transitions from one sound to another a nonlinear area transition is proposed which improve the transition in contrast to a purely linear transition of lossy and lossless models. The investigations show that the lossy tube model is advantageous compared to the lossless standard tube model for speech analysis and speech production.