Abstract

This paper describes novel and powerful variational EM algorithms for the segmental switching state space models used in speech applications, which are capable of capturing key internal (or hidden) dynamics of natural speech production. Hidden dynamic models (HDMs) have recently become a class of promising acoustic models to incorporate crucial speech-specific knowledge and overcome many inherent weaknesses of traditional HMMs. However, the lack of powerful and e±cient statistical learning algorithms is one of the main obstacles preventing them from being well studied and widely used. Since exact inference and learning are intractable, a variational approach is taken to develop effective approximate algorithms. We have implemented the segmental constraint crucial for modeling speech dynamics and present algorithms for recovering hidden speech dynamics and discrete speech units from acoustic data only. The effectiveness of the algorithms developed are verified by experiments on simulation and Switchboard speech data.