Estimating phone lengths for a diphone-based text-to-speech system for Arabic
Allan M. Ramsay, Hanady Mansour · Research Explorer (The University of Manchester) · 2008
We have described elsewhere a text-to-speech (TTS) system for Modern Standard Arabic which imposes a pitch contour on the output to indicate the force of the utterance (statement/query/command) and to mark emphasis (as specified by the use of non-canonical word orders). This TTS uses the diphone-based speech synthesiser Mbrola, for which you have to provide information about phone lengths. In the original implementation of the system we hand-coded this information. The current paper describes an approach to estimating phone lengths automatically by matching synthesised speech with human speech and adjusting the specification of individual phone lengths accordingly.