Improving speech synthesis by analysis by rule

This paper is concerned with details of the adaption process: a method of speech-analysis-by-synthesis-by-rule (`analysis-by-rule´) which provides the parameters of an underlying speech production model from an analysis of natural speech. This process is considered as a useful technique in its own right for analysing the acoustic realisations of a phonetic segment sequence. In particular, this paper addresses the question of acoustic variability and demonstrates that the analysis technique can separate predictable changes in segment realisation due to phonetic context from unpredictable changes due to variability in articulation, environment or analysis. In their original proposal Bridle and Ralls (1985) use a least-squares fitting technique to fit their SbR model to a corpus of natural speech to obtain the best data table values. This can be seen to confuse the two types of variability; whilst random variability can be captured as probability density functions for the data table values, the contextual variability requires changed to the rule-structure of the synthesis system