Doina Jitca

Improved Speech Synthesis Using Fuzzy Methods

The paper presents theoretical support for and describes the use of a fuzzy paradigm in implementing a TTS system for the Romanian language, employing a rule-based formant synthesizer. In the framework of classic TTS systems, we propose a new approach in order to improve formant trace computation, aiming at increasing synthetic speech perceptual quality. A fuzzy system is proposed for solving the problem of the phonemes that are prone to multi-definitions in rule-based speech synthesis. In the introductory section, we briefly present the background of the problem and our previous results in speech synthesis. In the second section, we deal with the problem of the context-dependent phonemes at the letter-to-sound module level of our TTS system. Then, we discuss the case of the phoneme /l/ and the solution adopted to define it for different contexts. A fuzzy system is associated with each parameter (denoted F1 and F2) to implement the results of the complete analysis of the phoneme /l/ behavior. The knowledge used in implementing the fuzzy module is acquired by natural speech analysis. In the third section, we exemplify the computation of the synthesis parameters F1 and F2 of the phoneme /l/ in the context of the two syllable sequences. The parameter values are contrasted with those obtained from the spectrogram analysis of the natural speech sequences. The last section presents the main conclusions and further research objectives.

The F0 contour modelling as functional accentual unit sequences

This paper presents an intonation description language based on the decomposition of complex intonational phrases (IP) into a tree of accentual units (AUs) and accentual unit group (AUGs) to which we have assigned functional labels at the communicative act level. A set of functional AU categories was defined and corresponding F0 contour patterns were described after performing an analysis over a Romanian speech corpus. Another important resource is the dictionary containing the descriptions of IPs/AUGs non-elementary melodic contours as functional unit sequences to which relative tonal coordinate sequences were assigned. We consider these functional labels suited to apply invariant meanings to different F0 contour units, structured into a functional unit hierarchy at the utterance level. We used the description language to perform flexible microprosodic descriptions of the text in a Romanian Text-to-Speech (TtS) system in order to control the F0 contour generation.

A prosodic control module for a Romanian TtS system, based on melodic contour dictionaries

This paper presents a prosodic control module implementation for a Romanian TtS system based on the F0 contour generation as a sequence of functional elementary melodic contours. The accentual units (AUs) are considered as elementary melodic contours because they contain elementary tonal contrasts to which labels with discourse signification may be assigned. Based on this idea, a set of AU F0 pattern categories was defined after performing an analysis over a Romanian speech corpus. The dictionary containing the parametric descriptions of the AU F0 pattern types in all possible lexical and phonetic contexts is an important resource in the prosodic control module. Another important resource is the dictionary containing the functional descriptions of the partial melodic contours at different intonational phrase (IP) or AU group (AUG) levels. After the intonational model and the label set presentation, we exemplify how our intonational description can be used to annotate the F0 contour. In the next section, the processing steps of the prosodic control module is presented.

An intonation prediction module for Romanian TTS system, as a prosodic tree generator

This paper presents a prosodic prediction module used by the Romanian TtS system in intonation synthesis. The module builds the input data for a prosodic control module that generates the F0 contour of the synthesized utterance. Both modules are based on a functional intonation model. The prediction module implies two main processing steps: identifying prosodic groups at different levels of a prosodic hierarchy and the selection of the melodic contours at each group level. Both processing steps use prosodic indications deduced from the text analysis. The first step is implemented by phrasing submodules that finally generates the utterance tree of the synthesized utterance. The second step refers to submodule that encodes the information that is relevant for the melodic contour selection from a melodic contour dictionary. The data output of the prediction module consists in the prosodic structure of the input text and the melodic contour descriptions at each prosodic units.

A review of several applications of FL in modeling and in 1D/2D signal processing

The paper is dedicated to the 50 years passed since the theory of fuzzy sets and fuzzy logic was established. This opportunity is used to analyze some research directions in the last decades and to suggest what directions may be fruitful to develop. The paper undertakes the task of reviewing some of our laboratory applications in 1D/2D signal processing. We are referring to the wave propagation fuzzy modeling, the use of fuzzy logic in formant detection and speech synthesis, in Parkinson disease diagnosis using handwriting text analysis and in processing grayscale medical images. Some results not already reported are also shown.

Information structure in Romanian utterances with contrast relations

The paper presents information packaging structures in Romanian utterances with the contrast relation, by decomposing them into hierarchies of embedded communicative units. At any level of the hierarchy, communicative units are structured by two or three functional constituents each of them having text and melodic contour. Communicative unit constituents are functional elements at the information structure level produced by their corresponding F0 patterns. The new account of information structure consists in working with two independent levels: the topic-focus and the referential information structures. Romanian sentences with contrast relation are discussed in the paper by presenting the influence of different discourse contexts on the communicative organization of the related utterances.

Aspects of a Romanian syntactic-prosodic interface for an intonation prediction module

This paper presents an improved version of a previously developed prosodic prediction module (PPM), used by a Romanian TtS (Text-to-Speech) system in intonation synthesis. The present PPM implements a method for setting prosodic indications, useful both in generating prosodic hierarchies and in melodic contour selection at each group level. The prosodic indications are derived from the lexical and morpho-syntactic analysis of the input text. The text analysis aims to generate a text structure by using only trees with one head and one or two branches, which describe local or global relations between words and/or group of words. This text hierarchy can be easily translated by the PPM into a prosodic hierarchy. Then, the elements of the relations are connected to different functional F0 patterns by selecting appropriate melodic contours, based on prosodicindications.

A Macro-Prosodic Indications for a Romanian TtS System Based on the Functional Intonational Model

This paper presents how macro-prosodic indications can be used within our TtS system in order to drive the prosody prediction module (PPM) in generating a target intonational contour for the synthesized utterance. The previous variant of the PPM generates melodic contour descriptions only using the implicit prosodic indications deduced from the text analysis. The explicit indications are edited in the input text window and refer to the group markers applied to a text to which can be assigned a focus attribute with an indication on its strength. It can be also assigned to a group an attribute of tonal prominence (related to the tonal level of the highest peak in the group). We called these indications the macro-prosodic indications. We also defined a microprosodic indication that refers to the pitch accent type attribute at the accentual unit level that is related to F0 contour pattern. The values of this attribute are taken from the Ro-ToBI label set. Both macro and microprosodic indications lead to elicit certain intonational variants at the speech synthesis output and also can improve the Romanian intonation contour understanding.

Prosodic phrases and contrast units in intonation

The paper explains the relation between prosodic phrases and Information Structure (IS) by decomposing phrases into hierarchies of embedded contrast/communicative units (CUs). At any level of hierarchies, CUs contains IS partitions supported by two contrasted functional constituents. The functional categories are defined by using a two level IS model. Topic-Focus and CU_predicate-CU_argument are the two independent levels of the proposed IS model. The paper presents different types of IS partition hierarchies for IPs and ips by using F0 contours of certain Romanian utterances. The IS hierarchy patterns are used in low level phrase identification within complex intonational phrases.

Intonation description language for the prosodic control in a TtS system for Romanian

This paper presents an intonation description language for the prosodic control implementation into a TtS system for Romanian. The prosodic description implies the decomposing of the F0 contours into of functional elementary melodic contours sequences. In our view the accentual units are considered as elementary melodic contours because they contain elementary tonal contrasts to which labels with communicative signification may be assigned. After the label set presentation from the third section, in the fourth section we exemplify how our description language could be used to annotate and generate the F0 contours.