Abstract
The human speech apparatus is a rich source of information and offers many cues in the speech signal due to its biomechanical constraints and physiological interdependencies. Coarticulation, a direct consequence of these speech production factors, is one of the main problems affecting the performance of speech systems. Incorporation of production knowledge could potentially benefit speech recognisers and synthesisers. Hand coded rules and scores derived from the phonological knowledge used by production oriented models of speech are simple and incomplete representations of the complex speech production process. Statistical models built from measurements of speech articulation fail to identify the cause of constraints. There is a need for building explanatory yet descriptive models of articulation for understanding and modelling the effects of coarticulation. This thesis aims at providing compact descriptive models of realistic speech articulation by identifying and capturing the essential characteristics of human articulators using measurements from electro-magnetic articulography. The constraints on articulators during speech production are identified in the form of critical, dependent and redundant roles using entirely statistical and data-driven methods. The critical role captures the maximally constrained target driven behaviour of an articulator. The dependent role models the partial constraints due to physiological interdependencies. The redundant role reflects the unconstrained behaviour of an articulator which is maximally prone to coarticulation. Statistical target models are also obtained as the by-product of the identified roles. The algorithm for identification of articulatory roles (and estimation of respective model distributions) for each phone is presented and the results are critically evaluated. The identified data-driven constraints obtained are compared with the well known and commonly used constraints derived from the IPA (International Phonetic Alphabet). The identified critical roles were not only in agreement with the place and manner descriptions of each phone but also provided a phoneme to phone transformation by capturing language and speaker specific behaviour of articulators. The models trained from the identified constraints fitted better to the phone distributions (40% improvement) . The evaluation of the proposed search procedure with respect to an exhaustive search for identification of roles demonstrated that the proposed approach performs equally well for much less computational load. Articulation models built in the planning stage using sparse yet efficient articulatory representations using standard trajectory generation techniques showed some potential in modelling articulatory behaviour. Plenty of scope exists for further developing models of articulation from the proposed framework.