Identifying prosodic prominence patterns for English text-to-speech synthesis

This thesis proposes to improve and enrich the expressiveness of English Textto-Speech (TTS) synthesis by identifying and generating natural patterns of prosodic prominence. In most state-of-the-art TTS systems the prediction from text of prosodic prominence relations between words in an utterance relies on features that very loosely account for the combined effects of syntax, semantics, word informativeness and salience, on prosodic prominence. To improve prosodic prominence prediction we first follow up the classic approach in which prosodic prominence patterns are flattened into binary sequences of pitch accented and pitch unaccented words. We propose and motivate statistic and syntactic dependency based features that are complementary to the most predictive features proposed in previous works on automatic pitch accent prediction and show their utility on both read and spontaneous speech. Different accentuation patterns can be associated to the same sentence. Such variability rises the question on how evaluating pitch accent predictors when more patterns are allowed. We carry out a study on prosodic symbols variability on a speech corpus where different speakers read the same text and propose an information-theoretic definition of optionality of symbolic prosodic events that leads to a novel evaluation metric in which prosodic variability is incorporated as a factor affecting prediction accuracy. We additionally propose a method to take advantage of the optionality of prosodic events in unit-selection speech synthesis. To better account for the tight links between the prosodic prominence of a word and the discourse/sentence context, part of this thesis goes beyond the accent/no-accent dichotomy and is devoted to a novel task, the automatic detection of contrast, where contrast is meant as a (Information Structure’s) relation that ties two words that explicitly contrast with each other. This task is mainly motivated by the fact that contrastive words tend to be prosodically marked with particularly prominent pitch accents. The identification of contrastive word pairs is achieved by combining lexical information, syntactic information (which mainly aims to identify the syntactic parallelism that often activates contrast) and semantic information (mainly drawn from the WordNet semantic lexicon), within a Support Vector Machines classifier. Once we have identified patterns of prosodic prominence we propose methods to incorporate such information in TTS synthesis and test its impact on synthetic speech naturalness trough some large scale perceptual experiments.