Articulatory Copy Synthesis Using Multiple Codebooks

A time domain synthmiser has been developed which uses nine parameters to control the shape of a model votarl tract.The mcdel, similar to Mermelstein's, was designed with the aid of magnetic resonance imaging (Nu-U) to nratch the vocal tract of a chosen speakert The parameters de ne the positions of the key articulators.from which the vocal tract outline is geometrically mnstructed.The am function is computed by superimposing a grid on the outline and using experimentally derived formulae relating say'ttal width to cross-sectional area.An acoustic-loaniculalory codebook was generated by randomly sampling the articulatory space spanned by the model.Copy synthesis was then performed by selecting a sequence of codebook entries with the aid of a dynamic programming algorithm.in mnjunction with a cost function that took into account both spectral and geometric information.The quality of the speech synthesised in this manner was restricted by the sparseness ofthe codebooltt The use ofa larger eodebook is proh riled due to the storage requirements and the computational e ort needed to search it.Clustering the codeboolr into regions in acoustic space.each de ned by a voiced diphone, allotted smaller codebooks to be used.Small'smle expansion of these sub-coddtooks permitted a series of small densely populated oodebooks to be generated.A complete utterance was then synthesised by using a di 'crent sub-codebook for each segment.Informal listening tests have revealed that speech quality was improved when synthesised using unis method.