Publications by Fiona Couper Kenney

The field of speech recognition research has been dominated by the Hidden Markov Model (HMM) in recent years. The HMM has known weaknesses, such as the strong “independence assumption” which presumes observations to be uncorrelated. New types of statistical modelling are now being investigated to overcome the weaknesses of HMMs. One such model is the Linear Dynamical Model (LDM), whose properties are more appropriate to speech. Modelling phone segments with LDMs gives fairly good classification and recognition scores, and this report explores possible extensions to a system using such models. Training only one model per phone cannot fully model variation that exists in speech, and perhaps training more than one model for some segments will improve accuracy scores. This is investigated here, and four methods for building two models instead of one for any phone are presented. Three of the methods produce significantly increased classification accuracy scores, compared to a set of single models.