Generating Synthetic Speech Prosody with Lazy Learning in Tree Structures

In: Proceedings of CoNLL-2000 and LLL-2000, pages 87-90, Lisbon, Portugal, 2000. Generating Synthetic Speech Prosody with Lazy Learning in Tree Structures Laurent Blin IRISA-ENSSAT F-22305 Lannion, France blin@enssat, fr Laurent Miclet IRISA-ENSSAT F-22305 Lannion, France miclet@enssat, fr Abstract We present ongoing work on prosody predic- tion for speech synthesis. This approach con- siders sentences as tree structures and infers the prosody from a corpus of such structures using machine learning techniques. The predic- tion is achieved from the prosody of the closest sentence of the corpus through tree similarity measurements, using either the nearest neigh- bour algorithm or an analogy-based approach. We introduce two different tree structure rep- resentations, the tree similarity metrics consid- ered, and then we discuss the different predic- tion methods. Experiments are currently under process to qualify this approach. 1 Introduction Natural prosody production remains a problem in speech synthesis systems. Several automatic prediction methods have already been tried for this, including decision trees (Ross, 1995), neu- ral networks (Traber, 1992), and HMMs (Jensen et al., 1994). The original aspect of our pre- diction approach is a tree structure representa- tion of sentences, and the use of tree similar- ity measurements to achieve the prosody pre- diction. We think that reasoning on a whole structure rather than on local features of a sen- tence should better reflect the many relations influencing the prosody. This approach is an attempt to achieve such a goal. The data used in this work is a part of the Boston University Radio (WBUR) News Cor- pus (Ostendorfet al., 1995). The prosodic infor- mation consists of ToBI labeling of accents and breaks (Silverman et al., 1992). The syntactic and part-of-speech informations were obtained from the part of the corpus processed in the Penn Treebank project (Marcus et al., 1993). We firstly describe the tree structures defined for this work, then present the tree metrics that we are using, and finally discuss how they are manipulated to achieve the prosody prediction. 2 Tree Structures So far we have considered two types of struc- tures in this work: a simple syntactic structure and a performance structure (Gee and Grosjean, 1983). Their comparison in use should provide some interesting knowledge about the usefulness or the limitations of the elements of information included in each one. 2.1 Syntactic Structure The syntactic structure considered is built ex- clusively from the syntactic parsing of the given sentences. This parsing, with the relative syn- tactic tags, constitute the backbone of the struc- ture. Below this structure lie the words of the sentence, with their part-of-speech tags. Addi- tional levels of nodes can be added deeper in the structure to represent the syllables of each word, and the phonemes of each syllable. The syntactic structure corresponding to the sentence "Hennessy will be a hard act to follow" is presented in Figure 1 as an example (the syl- lable level has been omitted for clarity). 2.2 Performance Structure The performance structure used in our approach is a combination of syntactic and …

Content truncated to 500 words