Speaker: Ganesh Sivaraman Seminar Title: Learning the Variability of Speech from Speech Production Abstract: Speech acoustic patterns vary significantly as a result of coarticulation and lenition processes that are shaped by segmental context or by performance factors such as rate of speech and degree of casualness/precision. Such variability offers serious challenges for the development of automatic speech recognition (ASR) systems that can perform well with minimal constraints. Articulatory phonology proposes that human speech can be decomposed into a constellation of articulatory gestures and it provides a unified framework for understanding how spatiotemporal changes in the pattern of underlying speech gestures lead to acoustic consequences that are typically reported as assimilations, insertions, deletions and substitutions. Motivated by this framework, movements of the articulators are expressed as a series of vocal tract constriction variables and articulatory gestures. Systems that estimate tract variables (TVs) from the speech signal are known as speech inversion systems. This talk focuses on speech inversion systems trained to estimate TVs from natural speech data. Tests on challenging scenarios of fast spoken speech provide encouraging results to use these systems to analyze the variability of speech. Estimated TV trajectories have also shown promising results on a small vocabulary speech recognition task in non-stationary noise conditions, thus motivating the application of speech production features to large vocabulary and multi-accent speech recognition.

• published: 11 Oct 2015
• views: 16