Antonio Castellanos

The EuTrans Spoken Language Translation System

The EuTransAll project aims at using example-based approaches for the automatic development of Machine Translation systems accepting text and speech input for limited-domain applications. During the first phase of the project, a speech-translation system that is based on the use of automatically learned subsequential transducers has been built. This paper contains a detailed and mostly self-contained overview of the transducer-learning algorithms and system architecture, along with a new approach for using categories representing words or short phrases in both input and output languages. Experimental results using this approach are reported for a task involving the recognition and translation of sentences in the hotel-receptioncommunication domain, with a vocabulary of 683 words in Spanish. Atranslation word-error rate of 1.97% is achieved in real-timefactor 2.7 on a Personal Computer.

An analysis of general acoustic-phonetic features for Spanish speech produced with the Lombard effect

Abstract A noisy environment usually degrades the intelligibility of a human speaker or the performance of a speech recognizer. Due to this noise, a phenomenon appears which is caused by the articulatory changes made by speakers in order to be more intelligible in the noisy environment: the Lombard effect. Over the last few years, special emphasis has been placed on analyzing and dealing with the Lombard effect within the framework of Automatic Speech Recognition. Thus, the first purpose of the work presented in this paper was to study the possible common tendencies of some acoustic features in different phonetic units for Lombard speech. Another goal was to study the influence of gender in the characterization of the above tendencies. Extensive statistical tests were carried out for each feature and each phonetic unit, using a large Spanish continuous speech corpus. The results reported here confirm the changes produced in Lombard speech with regard to normal speech. Nevertheless, some new tendencies have been observed from the outcome of the statistical tests.

Text and speech translation by means of subsequential transducers

The full paper explores the possibility of using Subsequential Transducers (SST), a finite state model, in limited domain translation tasks, both for text and speech input. A distinctive advantage of SSTs is that they can be efficiently learned from sets of input-output examples by means of OSTIA, the Onward Subsequential Transducer Inference Algorithm (Oncina et al. 1993). In this work a technique is proposed to increase the performance of OSTIA by reducing the asynchrony between the input and output sentences, the use of error correcting parsing to increase the robustness of the models is explored, and an integrated architecture for speech input translation by means of SSTs is described.

Language understanding and subsequential transducer learning

Language Understanding can be considered as the realization of a mapping from sentences of a natural language into a description of their meaning in an appropriate formal language. Under this viewpoint, the application of the Onward Subsequential Transducer Inference Algorithm (OSTIA) to Language Understanding is considered. The basic version of OSTIA is reviewed and a new version is presented in which syntactic restrictions of the domain and/or range of the target transduction can effectively be taken into account. For experimentation purposes, a task proposed by Feldman, Lakoff, Stolcke and Weber (1990) (International Computer Science Institute, Berkley, California) for assessing the capabilities of language learning and understanding systems has been adopted and three semantic coding schemes have been defined for this task with different sources of difficulty. In all cases the basic version of OSTIA has proved consistently to be able to learn very compact and accurate transducers from relativly small training sets of input?output examples of the task. Moreover, if the input sentences are corrupted with syntactic incorrectness or errors, the new version of OSTIA still provides understandable results that only degrade in a gradual and natural way.

Some results with a trainable speech translation and understanding system

The problems of limited-domain spoken language translation and understanding are considered. A standard continuous speech recognizer is extended for using automatically learnt finite-state transducers as translation models. Understanding is considered as a particular case of translation where the target language is a formal language. From the different approaches compared, the best results are obtained with a fully integrated approach, in which the input language acoustic and lexical models, and (N-gram) language models of input and output languages, are embedded into the learnt transducers. Optimal search through this global network obtains the best translation for a given input acoustic signal.

Approximate Learning of Random Subsequential Transducers

In this work, approximate inference of random partial Subsequential Transducers (STs) is addressed. Accessibility and distinguishability of a ST are defined and used to bound the maximum length of samples which are going to form representative sets for target STs. From these representative sets, the sample density required to obtain good approximate STs has been investigated. Dependency of the sample density on the number of states and on the accessibility and distinguishability of the target STs has been evaluated. As a general result, a decrease of the sample density has been found as these parameters increase, suggesting that accessibility and distinguishability are parameters as important as the number of states to evaluate learnability of STs.