Abstract. This paper describes an approach to improving synthesized speech quality for voices created by using an audiobook database. The data consist of a large amount of read speech by one speaker, which we matched with the corresponding book texts. The main problems with such a database are the following. First, the recordings were made at different times under different acoustic conditions, and the speaker reads the text with a variety of intonations and accents, which leads to very high voice parameter variability. Second, automatic techniques for sound file labeling make more errors due to the large variability of the database, especially as there can be mismatches between the text and the corresponding sound files. These problems dramatically affect speech synthesis quality, so a robust method for solving them is vital for voices created using audiobooks. The approach described in the paper is based on statistical models of voice parameters and special algorithms of speech element concatenation and modification. Listening tests show that it strongly improves synthesized speech quality.

Abstract. In this paper we present a system for automatically predicting prosodic breaks in synthesized speech using the Random Forests classifier. In our experiments the classifier is trained on a large dataset consisting of audiobooks, which is automatically labeled with phone, word, and pause labels. To provide part of speech (POS) tags in the text, a rule-based POS tagger is used. We use crossvalidation in order to be able to examine not only the results for a specific subset of data but also the systems reliability across the dataset. The experimental results demonstrate that the system shows good and consistent results on the audiobook database; the results are poorer and less robust on a smaller database of read speech even though part of that database was labeled manually.