Automatic speech recognition and text-to-speech systems have modules that depend on the language and, while there are many public resources for some languages (e.g. English and Japanese), the resources for Brazilian Portuguese (BP) are still limited. Another aspect is that for many tasks the current speech recognition system error rate is still high, when compared
to that obtained by humans. Thus, despite the success of hidden Markov models (HMM), it
is necessary to investigate new methods. This work has these two facts as motivation and is divided into two parts. The first part describes the resources and free tools developed for BP speech recognition and synthesis, consisting of text and audio databases, phonetic dictionary, grapheme-to-phone converter, syllabification module, language and acoustic models. All of them are publicly available and, together with a proposed application programming interface,
have been used for the development of several new real-time applications, including a speech module for the OpenOffice suite. Performance tests are presented for evaluating the developed systems. The resources make easier the adoption of BP speech technologies by other academic
groups, developers and industry. The second part of this work presents a new method for
rescoring the recognition result obtained via HMMs, with the result being organized as a lattice. More specifically, the system uses discriminative classifiers that aim at reducing the confusability between pairs of phones. For each of these binary problems, automatic feature selection techniques are used to choose the proper parametric representation for the specific problem.
|
