Céu Viana

DIXI --- A Generic Text-to-Speech System for European Portuguese

This paper describes a new generic text-to-speech synthesis system, developed in the scope of the Tecnovoz Project. Although it was primarily targeted at speech synthesis in European Portuguese, its modular architecture and flexible components allows its use for different languages. We also provide a survey on the development of the language resources needed by the TTS.

Language and variety verification on broadcast news for Portuguese

This paper describes a language/accent verification system for Portuguese, that explores different type of properties: acoustic, phonotactic and prosodic. The two-stage system is designed to be used as a pre-processing module for the Portuguese Automatic Speech Recognition (ASR) system developed at INESC-ID. As the ASR system is applied everyday to transcribe the evening news from a Portuguese public TV channel, the presence of other languages (mainly English) and other varieties of Portuguese is very likely. In the first stage, for each automatically detected speaker, the system verifies if the spoken language is Portuguese, as opposed to nine other languages - English, Belgian Dutch, Croatian, Czech, Galician, Greek, Hungarian, Sloven and Slovak. The identified Portuguese speakers are then fed to the second stage which aims at identifying the Portuguese variety: European, Brazilian or African Portuguese from five countries. The identification results are then used either to mark the speech data as untranscribable or forward it to the European Portuguese ASR system, or a system tuned for other languages or varieties. The language verification system achieved an equal error rate for European Portuguese of 2.5%. In terms of variety identification, the overall rate of correct identification was 83.9%, when considering only the three broad varieties, and the best results were obtained for Brazilian Portuguese, also the variety that proved easiest to identify in perceptual experiments. The identification rate between African varieties themselves is relatively low, a fact that was also observed in the perceptual experiments.

Prosodic Phrasing: Machine and Human Evaluation

This paper describes a set of experiments aiming at the construction and evaluation of a new phrasing module for European Portuguese text-to-speech synthesis, using classification and regression trees learned from hand-labelled texts. Using the assessment criteria of matching boundary predictions against the corresponding labelled ones, the best solution achieves an overall performance of 91.9%, with 86.3% of correctly assigned breaks and 4.3% of false insertions. Although in absolute terms such scores may be considered surprisingly good given the size of the training set, the total number of exact matches at the sentence level is much lower (22%). This suggested a more formal experiment to test the acceptability of the predicted phrasing in the judgement of human evaluators. As the model was not trained on a labelled speech corpus but on hand-labelled texts, the reference phrasing needed also to be assessed. The evaluation experiment involved 90 participants who were asked to grade both the automatic and the reference phrasings, and also to express their opinion on where the breaks should be placed. As expected, the results showed a large variability among the subjects in their acceptance of a specific sentence partition, and criteria had to be defined to summarise the data from the different evaluators. With the adopted criteria, the performance of the automatic assignment procedure at the sentence level is better rated by human evaluators than by simple matching with the reference corpus (78% vs. 22%, respectively).

An Electronic Assistant for Poetry Writing

The ultimate goal of the poetry assistant currently under development in our lab is an application to be used either as a poetry game or as a teaching tool for both poetry and grammar, including the complex relationships between sound and meaning. Until now we focused on the automatic classification of poems and the suggestion of the ending word for a verse. The classification module is based on poetic concepts that take into account structure and metrics. The prediction module uses several criteria to select the ending word: the structural constraints of the poem, the grammatical category of the words, and the statistical language models obtained from a text corpus.

From Portuguese to Mirandese: fast porting of a letter-to-sound module using FSTs

This paper describes our efforts in porting our letter-to-sound module from European Portuguese to Mirandese, the second official language in Portugal. We describe the rule formalism and the composition of the various transducers involved in the letter-to-sound conversion. We propose a set of extra SAMPA symbols to be used in the phonetic transcription of Mirandese, and we briefly cover the set of rules and results obtained for the two languages. Although at a very preliminary stage, we also describe our efforts at building a waveform generation module also based on finite state transducers. The use of finite state transducers allowed a very flexible and modular framework for deriving and testing new rule sets. Our experience led us to believe that letter-to-sound modules could be helpful tools for researchers involved in the establishment of orthographic conventions for lesser spoken languages.

Portuguese variety identification on broadcast news

This paper describes an accent identification system for Portuguese, that explores different type of properties: acoustic, phono tactic and prosodic. The system is designed to be used as a pre-processing module for the Portuguese automatic speech recognition system developed at INESC-ID. In terms of variety identification, the overall rate of correct identification is 69.0% if all 7 varieties are considered, and the best results are obtained for Brazilian Portuguese, also the variety that proved easiest to identify in perceptual experiments. When distinguishing between European, Brazilian and African Portuguese, the identification rate goes up to 94.7%. The fact that the prosodic system alone can achieve an identification rate of 77% is also worth investigating.