Compression of prosody for speech modification in synthesis

Methods of compressing the prosodic information about a speaker´s delivery-pitch, duration, and amplitude - are described. The objective is to use available or extracted knowledge of the spoken text along with the prosodic information to synthesize speech from a suitable inventory of stored basic units. Techniques for compressing pitch and amplitude of speech units using transform coding are investigated. Discrete cosine and sine transforms are found to be effective in compressing pitch and amplitude information respectively. In order to generate speech with these coded pitch and amplitude contours, the prosodic features of speech units stored in an inventory are modified using a method that was proposed to perform speech modification for concatenative synthesis. In this method, the stored speech unit is processed with a suitably shaped time-varying prefilter, whose parameters are chosen to have a low sensitivity to pitch changes. The filtered signal is modified according to the required change in its prosodic structure, and then applied to the inverse of the prefilter. Examples of application of the proposed representation and modification of prosody to a variety of speech units are presented.