In this paper we address the issue of pronunciation modeling for conversational speech synthesis. We experiment with two different HMM topologies (fully connected state model and forward connected state model) for sub-phonetic modeling to capture the deletion and insertion of sub-phonetic states during speech production process. We show that the experimented HMM topologies have higher log likelihood than the traditional 5-state sequential model. We also study the first and second mentions of content words and their influence on the pronunciation variation. Finally we report phone recognition experiments using the modified HMM topologies.

/pdf/sub-phonetic-modeling-for-capturing-pronunciation-variations-1u5u6ldw9r.pdf

Sub-Phonetic Modeling For Capturing Pronunciation Variations For Conversational Speech Synthesis

The Pairwise Variability Index (PVI), a measure of how much unit-to-unit variation there is in speech, has been used as a correlate of rhythmic impressions such as ‘syllable-timed’ and ‘stress-timed’. Grabe and Low (2002) included Estonian among a number of languages for which they calculated the durational PVI for vowels and for intervocalic intervals, but other than that Estonian rhythm has not been studied within recent approaches to rhythm calculation. The pilot experiment reported in this paper compares the use of various speech units for the characterisation of Estonian speech rhythm. It is concluded that the durational PVI of the syllable and of the foot are more appropriate for capturing the rhythmic complexity of Estonian, and might provide a subtle tool for characterising languages in general.

Estonian rhythm and the Pairwise Variability Index

While text-to-speech has long been centered on the production of an intelligible message of good quality, interest has recently shifted to the generation of more natural and expressive speech. This comes as an answer to the widespread criticism stating that current speech synthesizers lack fundamental human components. This thesis tackles that issue by considering three fundamental stages of HMM-based speech synthesis: the phonetic and prosodic annotations of the training corpus and their automatic alignment with the speech signal. We first propose a systematic step-by-step study of HMM-based phonetic alignment in which the models are directly trained on the corpus to align. Based on a detailed analysis of the errors made by this technique, we developed three fully-automatic improvement methods which are shown to significantly improve the alignment of highly variable and expressive corpora. We then define a two-level prosodic annotation of expressive corpora, describing accentual patterns and changes in speaking style. The integration of this manual annotation in the synthesis of sports commentaries positively impacts the naturalness of the expressivity. We also present an automatic annotator of accentual patterns in French and show that its integration in synthesis contributes to the naturalness of the voice. Finally, our study points out that the choice for phonetic variants in French is influenced by the speaking style and that their consideration in the synthesis of sports commentaries improves the naturalness of the message. It indicates that phonetic changes should be considered, both at training and synthesis stages.

Expressive speech synthesis : research and system design with hidden Markov models

In this paper, pronunciation variation in Swedish due to speaking style and speech rate is discussed. A tentative rule system for segment-level reduction is currently being evaluated by letting subjects assess the naturalness of synthetic speech generated from canonical transcriptions and transcriptions reduced by the system, respectively. Results from experiments using short sentences with explicit control over the rules applied have shown that reduced forms are preferred at high speech rates (rates above the synthesis default rate), while there is no significant bias in preference between canonical and reduced forms at the synthesis default speech rate. Presently, longer and less controlled passages of synthetic speech are being evaluated using the same experiment set-up. Using text passages of varying degree of formality, this experiment allows for testing the effects of text formality on perceived naturalness.

/pdf/evaluating-rules-for-phonological-reduction-in-swedish-3lt1n52r6l.pdf

Evaluating rules for phonological reduction in Swedish

Analytic knowledge about the systematic variation in a language has an important place in the description of the language. Such knowledge is interesting e.g. in the language teaching domain, as a b ...

/pdf/modelling-phone-level-pronunciation-in-discourse-context-2o4rldya0i.pdf

Modelling Phone-Level Pronunciation in Discourse Context

In this paper, the importance of pronunciation varia- tion modelling is discussed. As a first step in devel- oping a model of Swedish pronunciation variation due to speaking style and speech rate, a tentative reduc- tion rule system has been developed. An assessment experiment testing the impact of phonological reduc- tion, as defined by this system, on the perceived natu- ralness of speech synthesis was conducted. Canonical and reduced synthetic speech stimuli with three differ- ent speech rates were presented to na¨ ive subjects. The reduced pronunciations were significantly more often perceived as more natural than the canonical pronun- ciations at the higher speech rates, while there was no significant general difference in perceived naturalness depending on reduction level for the lowest rate. The dependence on speech rate for perceived naturalness was significant. A possible cause for some observed differences in perceived naturalness depending on the nature of specific stimuli is discussed. For a general description of speaking style dependent pronunciation variation in Swedish, both phonological and phonetic level rules will have to be developed. For this purpose, data-driven methods will be used on an- notated spontaneous speech corpora. The focus will be on general aspects of pronunciation variation, rather than on variation due to dialect or individual factors. As a starting point, a tentative knowledge-based phonological level rule system has been developed. The purpose of this rule system is to form a base from which a more elaborate rule system can be built. The rule system has some empirical grounds, since it is based partly on empirical results reported by Garding (1). The result of applying the rule system to canonical (maximally detailed) transcriptions can be compared to what can be observed in spontaneous speech cor- pora and the rules updated, if systematic deviations are found. The phonological rules can also serve as a skeleton to which more detailed phonetic rules can be added. In this paper, an assessment experiment testing the impact on the perceived naturalness of speech synthesis of the tentative rule system is reported. The specific questions asked are: 1) What is the impact on the perceived naturalness of synthesis output of applying the rule system to the input transcriptions, 2) What is the correlation between the perceived naturalness of the rule-processed stimuli and synthesis speech rate and 3) Are there differences in perceived naturalness depending on the rules applied and, if so, how can these differences be explained? 1.1 BACKGROUND

/pdf/phonological-reduction-in-swedish-11mwd82lbu.pdf

Per-Anders Jande

Papers

Phonological Reduction in Swedish