This paper proposes a new scheme for concatenative speech synthesis to improve the speech segment selection procedure. The proposed scheme selects a segment sequence for concatenation by minimizing acoustic distortions between the selected segment and the desired spectrum for the target without the use of heuristics. Four types of distortion, a) the spectral prototypicality of a segment, b) the spectral difference between the source and target contexts, c) the degradation resulting from concatenation of phonemes, and d) the acoustic discontinuity between the concatenated segments, are formulated as acoustic quantities, and used as measures for minimization. A search method for selecting segments from a large speech database is also descrided. In this method, a three-step optimization using dynamic programming is used to minimize the four types of distortion. A perceptual test shows that this proposed segment selection method with minimum distortion criteria produces high quality synthesized speech, and that contextual spectral difference and acoustic discontinuity at the segment boundary are important measures for improving the quality.

This paper presents a new method for statistical modelling of prosody control in speech synthesis. The proposed method, which is referred to as Constrained Tree Regression (CTR), can make suitable representation of complex effects of control factors for prosody with a moderate amount of learning data. It is based on recursive splits of predictor variable spaces and partial imposition of constraints of linear independence among predictor variables. It incorporates both linear and tree regressions with categorical predictor variables, which have been conventionally used for prosody control, and extends them to more general models. In addition, a hierarchical error function is presented to consider hierarchical structure in prosody control. This new method is applied to modelling of speech segmental duration. Experimental results show that better duration models are obtained by using the proposed regression method compared with linear and tree regressions using the same number of free parameters. It is also shown that the hierarchical structure of phoneme and syllable durations can be represented efficiently using the hierarchical error function.

This paper describes an application of a speech translation system to another task/domain in the real-world by using developmental data collected from real-world interactions. The total cost for this task-alteration was calculated to be 9 Person-Month. The newly applied system was also evaluated by using speech data collected from real-world interactions. For real-world speech having a machine-friendly speaking style, the newly applied system could recognize typical sentences with a word accuracy of 90% or better. We also found that, concerning the overall speech translation performance, the system could translate about 80% of the input Japanese speech into acceptable English sentences.

We propose new language models to represent phrasal structures by patterns extracted from parse trees. First, modified word trigram models are proposed. They are extracted from sentences analyzed by the preprocessing of the parser with knowledge. Since sentences are analyzed to create sub-trees of a few words, these trigram models can represent relations among a few neighbor words more strongly than conventional word trigram models. Second, word pattern models are used on these modified word trigram models. The word patterns are extracted from parse trees and can represent phrasal structures and much longer word-dependency than trigram models. Experimental results show that modified trigram models are more effective than traditional trigram models and that pattern models attain slight improvements over modified trigram models. Furthermore, additional experiments show that pattern models are more effective for long sentences.

In this paper, we quantitively analyzed speech data in seven different styles to make natural Japanese conversational speech synthesis. Three reading styles were produced at different speeds (slow, normal and fast), and four speaking styles were produced by enacting conversation in different situations (free, hurried, angry and polite). To clarify the differences in prosodic characteristics between conversational speech and read speech, means and standard deviations of vowel duration, vowel amplitude and fundamental frequency (F0) were analyzed. We found large variation in these prosodic parameters. To look more precisely at the segmental duration and segmental amplitude differences between conversational speech and read speech, control rules of prosodic parameters in reading styles were applied to conversational speech. F0 contours of different speaking styles are superposed by normalizing the segmental duration. The differences between estimated values and actual values were analyzed. Large differences were found at sentence final and key (focused) phrases. Sentence final positions showed lengthening of segmental vowel duration and increased segmental vowel amplitude. Key phrase positions featured raising F0.