| This thesis focuses on exploring the role of speech production in automatic speech recognition from a communication system perspective. Specifically, I have developed a generalized smoothness criterion (GSC) for a talker-independent acoustic-to-articulatory inversion, which estimates speech production/articulation features from the speech signal of any arbitrary talker. GSC requires parallel articulatory and acoustic data from a single subject only (exemplar) and this exemplar need not be any of the talkers. Using both theoretical analysis and experimental evaluation, it is shown that the estimated articulatory features provide recognition benefit when used as additional features in an automatic speech recognizer. As we require a single exemplar for the acoustic-to-articulatory inversion, we overcome the need for the articulatory data from multiple subjects during inversion. Thus, we demonstrate a feasible way to utilize production-oriented features for speech recognition in a data-driven manner. Due to the concept of exemplar, the production-oriented features and, hence, the speech recognition become exemplar-dependent. Preliminary recognition results with different talker-exemplar combinations show that the recognition benefit due to the estimated articulatory feature is greater when the talker's and exemplar's speaking styles are matched, indicating that the proposed exemplar-dependent recognition approach has potential to explain the variability in recognition across human listeners. |
