Evaluation of an improved method for automatic extraction of model parameters from fundamental frequency contours of speech

Abstract

The authors have already presented a method for automatic extraction of accent and phrase commands of a model from a given F0 contour of speech. This paper describes improvements introduced to cope with difficulties encountered by the previous method, especially in connection with the extraction of accent commands, and reports the results of experiments conducted for the evaluation of the current method using two sets of speech materials differing in sentence length and syntactic complexity. It is shown that the method works quite well for the majority of utterances tested. Analysis of performance in terms of misses and false insertions of commands indicates that the performance is slightly better for shorter utterances, and that most of the errors are related to commands of smaller magnitude/amplitude, suggesting that their effects on the perception of naturalness of prosody are of minor importance.