Vera Demberg Livres

Grapheme-to-phoneme conversion (g2p) is a core component of any text-to-speech system. This book discusses how adding simple syllabification and stress assignment constraints, namely 'one nucleus per syllable' and 'one main stress per word', to a joint n-gram model for g2p conversion leads to a dramatic improvement in conversion accuracy. The model is evaluated on German, English and French, and compares well to state-of-the-art approaches. Secondly, the benefit to be gained from morphological preprocessing for g2p conversion is assessed. While morphological information has been incorporated in some past systems, its contribution has never been quantitatively assessed for German. We compare the relevance of morphological preprocessing with respect to the morphological segmentation method, training set size, the g2p conversion algorithm, and two languages, English and German. This work is particularly valuabel for professionals working with speech syntesis systems.

In spoken dialogue systems (e.g. for flight information or restaurant recommendation) it is difficult to present a large number of alternative options because enumerating all available options would take very long, and the user would not be able to remember the details. This work proposes to tackle this problem by identifing compelling options based on a model of user preferences, and presenting tradeoffs between alternative options explicitly. Multiple attractive options are structured such that the user can gradually refine her request to find the optimal tradeoff. Evaluation shows that the approach presents complex tradeoffs understandably, increases overall user satisfaction, and significantly improves the user's overview of the available options. Moreover, the results suggest that presenting users with a brief summary of the irrelevant options increases users' confidence in having heard about all relevant options. The strategy for information presentation proposed in this book should be directly useful to professionals designing dialogue systems, and can inform other related tasks such as the presentation of information in question answering or web search.