University of Patras
Μοντελοποίηση και ψηφιακή επεξεργασία προσωδιακών φαινομένων της ελληνικής γλώσσας με εφαρμογή στην σύνθεση ομιλίας
Abstract
dc:descriptionIn this thesis we cope with the task of studying and modeling prosodic phenomena encountered in Greek language with applications to the task of speech synthesis from text. For the achievement of this particular task, spoken corpora with various levels of morphosyntactical and linguistic representation as well as tools for their processing were constructed. With the success of concatenative speech synthesis as well as the shift from rule-based modeling of prosody to formal multivariate models that require the associated parameters to be statistically derived from a training corpus research in the area of synthetic speech production has come to rely on models produced from data driven methods. In particular concatenative speech synthesis heavily relies on the existence of an automatic selection procedure of candidate units from a recorded speech database. Furthermore construction of reliable data-driven prosodic event prediction models requires a well-designed annotated speech corpus and greatly depends on its quantity domain coverage as well as annotation consistency. As a result data collection and design requirements describing a prosodic corpus have a direct impact on the resulted synthetic speech of a Text to-Speech (TtS) system. However apart from the minimal requirement of having enough data for the training and testing phase of a statistical model estimation procedure there are no standard rules for corpus design and collection. Construction and annotation of the Greek prosodic speech data, the WCL-1 prosodic database took into account most of the phenomena that constitute the prosodic patterns of Greek language. Due to the fact that Greek is a minority language only restricted research has been conducted on the prosodical aspects of speech communication and therefore there is a scarcity of annotated speech data and tools for rapid database development. In the framework of this endeavor prosodic characteristics of spoken Greek were defined labeled and stored so that relations between intonational phenomena and textual as well as linguistic features could be first extracted and then modeled with the application of machine learning techniques. For the task of coding the emerged prosodic phenomena of our recorded utterances we have followed the work of Pierrehumbert for American English (Pierrehumbert 1981: 985-995). Furthermore our prosodic framework contains segmentation of the data on phonemic level. Several tools were developed in order to perform various tasks such as corpus creation recording and labeling. The data of the prosodic corpus were processed in the following order: (a) selection of a representative text corpus including the most common prosodic phenomena of the Greek language, (b) selection of the speaker and speaking style for the text corpus recording session, (c) phonetic segmentation and labeling, and (d) annotation of the tones and prosodic boundaries with the utilization of Tones and Break Indices (ToBI) annotation convention adopted for Greek. Corpus based synthesis has turned the attention of researchers in deriving accent placement and phrasing rules for text to-speech systems from large labeled corpora most recently attempts have been made to use self-organizing procedures to compute prosodic specification rules automatically from such data. The most commonly used learning techniques are Hidden Markov models, neural networks classification and regression trees transformational rule-based learning and Bayesian. A large amount of intonational variations were taken into account that proved to be useful for prosodic training. Specifically we focused in the task of predicting break indices and pitch accent labels from textual features.
Degree
thesis:*- Grantor dc:publisher
- University of Patras
- Year dc:date
- 2007
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Zervas, Panagiotis
- Ζέρβας, Παναγιώτης
Subjects
dc:subject × 14Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/26654