Multimodal Speech Synthesis

A Computational Model of Unsupervised Speech Segmentation for Correspondence Learning

“…For example, in our corpus (Schweitzer et al 2004) fewer than 20% of the 1 Related work by Xanthos (Goldsmith and Xanthos 2009) explores a range of methods for establishing whether it is possible to automatically infer whether segments in a data sample are vowels or consonants (in addition to examining vowel harmony and phonotactic induction). Segmentation, however, is not the focus.…”

“…We also use the IMS unit selection corpus (Schweitzer et al 2004), a corpus of German speech, recorded by a professional male and a professional female speaker, and sampled at 16,000 Hz. 13-dimensional MFCCs were computed for the 2776 original speech files of the male speaker-part by means of the MATLAB Auditory Toolbox (Slaney 1998).…”

A dynamical model for generating prosodic structure

“…The prosody-syntax interface has been a theme of great interest in the scientific community, as illustrated by the proposals of several algorithms that explain part of the variance of both prosodic constituency and prominence [1,6,8,9]. All these algorithms have been integrated into text-to-speech synthesis systems in order to automatically generate the prosodic information necessary to produce a natural-sounding speech.…”

“…Three depths of syntactic analysis prior to the obtention of prosodic structure can be identified in these models. Some use a comprehensive parser to analyse the sentences [9], others use a partial syntactic analysis [1], and the last ones use a minimal amount of syntactic information [6,8]. All the algorithms use a set of heuristic rules to obtain prosodic constituents of similar size.…”