Section: Research Program

Speech to Speech Translation and Language Modeling

Speech-to-Speech Translation aims at translating a source speech signal into a target speech signal. A sequential way to address this problem is to first translate a text to another one. And after, we can connect a speech recognition system at the input and a text to speech synthesis system at the output. Several ways to address this issue exist. The concept used in our group is to let the computer learning from a parallel text all the associations between source and target units. A unit could be a word or a phrase. In the early 1990s [80] proposes five statistical translation models which became inescapable in our community. The basic idea of the model 1 is to consider that any word of the target language could be a potential translation of any source word. The problem is then to estimate the distribution probability of a target word given a source one. The translation problem is similar to the speech recognition one. Indeed, we have to seek the best foreign sentence given a source one. This one is obtained by decoding a lattice translation in which a language and translation models are used. Several issues have to be supported in machine translation as described below.

The first translation systems identify one-to-one associations between words of target and source languages. This is still necessary in the present machine translation systems. In our group we develop a new concept to learn the translation table. This approach is based on computing all the inter-lingual triggers inside a parallel corpus. This leads to a pertinent translation table [95] . Obviously, this is not sufficient in order to make a realistic translation because, with this approach, one word is always translated into one word. In fact, it is possible to express the same idea in two languages by using different numbers of words. Thus, a more general one-to-one alignment has to be achieved.

The human translation is a very complex process which is not only word-based. A number of research groups developed phrase-based systems which are different from the baseline IBM's model in training. These methods deal with linguistic units which consists in more than one word. The model supporting phrase-based machine translation uses reordering concept and additional feature functions. In order to retrieve phrases, several approaches have been proposed in the litterature. Most of them require word-based alignments. For example, Och and al. [96] collected all phrase pairs that were consistent with the word alignment provided by Brown's models.

We developed a phrase based algorithm which is based on finding first an adequate list of phrases. Then, we find out the best corresponding translations by using our concept of inter-lingual triggers. A list of the best translations of sequences is then selected by using simulated annealing algorithm.

The international community uses either PHARAOH [93] or MOSES [92] based on a beam search algorithm. In our group we started decoding by PHARAOH but we moved recently to MOSES.