A maximum likelihood continuity mapping approach to discovering the intrinsic structure of speech articulation manifolds

Over the past few decades, significant improvements have been made in our ability to recover articulation from the acoustics of speech. Approaches have evolved from first principle models based on erroneous, simplifying assumptions to plausible, physiologically based data driven methods. One of these methods, called maximum likelihood continuity mapping, is built on the premise that output realizations close together in a high-dimensional observation space had to have been generated by positions close together in a low-dimensional latent space. This model is consistent with our understanding of how speech is produced; low-dimensional articulatory configurations in close proximity to each other help generate the temporally connected acoustics we hear.;In this thesis, maximum likelihood continuity mapping is used to discover the intrinsic parameters and, reduce the dimensionality of two speaker-dependent articulation manifolds from the MOCHA-TIMIT data set: the articulatory measurements of female speaker fsew0 and male speaker msak0. The reduced manifolds, called canonical articulation manifolds, are then compared to the manifolds of respective speakers to analyze the performance of maximum likelihood continuity mapping in modeling speech. Two areas are addressed. First, a direct comparison between canonical articulation manifolds and manifolds of articulatory measurements tells us how well maximum likelihood continuity mapping models articulation. Second, canonical articulation manifolds are compared to estimated articulation manifolds derived from the acoustic signal. This comparison tells us how effective maximum likelihood continuity mapping is in dealing with the one-to-many mappings encountered in recovering articulation from acoustics.;A key element in the analysis is the use of context to recover articulatory paths. Context is used to find the intrinsic dimensionality and cutoff frequency of articulation manifolds, to recover a low-dimensional canonical manifold from speech acoustics, to analyze the existence of and resolve one-to-many mappings found in the acoustic-to-articulation transformation, and to provide evidence that the use of multiple acoustic frames to recover one articulatory configuration improves the estimate by reducing the ambiguity of one-to-many mappings.