Variable-length unit selection using LSA-based syntactic structure cost

Abstract

The paper introduces a variable-length unit selection method for concatenative speech synthesis based on a syntactic structure based on latent semantic analysis (LSA). First, a probabilistic context free grammar (PCFG) based parser is used to construct the syntactic structure of the input text sentence. Second, the synthesizer selects the candidate units for each node of the syntactic structure. LSA is then adopted to estimate the syntactic cost between the target unit and the candidate units in the database. Finally, the concatenation of units with minimum cost is selected using a dynamic programming algorithm. Experimental results show that variable-length unit selection based on syntactic structure outperforms the synthesizer that does not consider syntactic structure. Also, the LSA-based syntactic cost provides a better estimation of substitution cost than that calculated only from acoustic features.