Catálogo de publicaciones - libros

In this work we show a performance improvement of our system by taking into account the weights of the mixture of Gaussians of the Hidden Markov Model. Furthermore and independently tunning of each of the phoneme Support Vector Machine (SVM) parameters is performed. In our system the user utters a pass phrase and the phoneme waveform segments are found using the Automatic Speech Recognition Technology. Given the speech model and the phoneme information in the segments, a set of features are created to train an SVM that could generate a cryptographic key. Applying our method to a set of 10, 20, and 30 speakers from the YOHO database, the results show a good improvement compared with our last configuration, improving the robustness in the generation of the cryptographic key.

In this paper, a hybrid language model which combines a word-based n-gram and a category-based Stochastic Context-Free Grammar (SCFG) is evaluated for training data sets of increasing size. Different estimation algorithms for learning SCFGs in General Format and in Chomsky Normal Form are considered. Experiments on the UPenn Treebank corpus are reported. These experiments have been carried out in terms of the test set perplexity and the word error rate in a speech recognition experiment.

The purpose of this work is to show the results obtained when the latest technological advances in the area of Automatic Speech Recognition (ASR) are applied to the Western-Huastec Náhuatl and Huastec languages. Western-Huastec Náhuatl and Huastec are not only native (indigenous) languages in México, but also minority languages, and people who speak these languages usually are analphabetic. A speech database was created by recording the voice of native speaker when reading a set of documents used for native bilingual primary school in the official mexican state education system. A pronunciation dictionary was created for each language. A continuous Hidden Markov Models (HMM) were used for acoustical modeling, and bigrams were used for language Modeling. A Viterbi decoder was used for recognition. The word error rate of this task is below 8.621% for Western-Huastec Náhuatl language and 10.154% for Huastec language.

The first pattern recognition approaches to machine translation were based on single-word models. However, these models present an important deficiency; they do not take contextual information into account for the translation decision. The phrase-based approach consists in translating a multiword source sequence into a multiword target sequence, instead of a single source word into a single target word. We present different methods to train the parameters of this kind of model. In the evaluation phase of this approach, we obtained interesting results in comparison with other statistical models.

Segmentation of bilingual text corpora is a very important issue to deal with in machine translation. In this paper we present a new method to perform bilingual segmentation of a parallel corpus, SPBalign , which is based on phrase-based statistical translation models. The new technique proposed here is compared with other two existing techniques, which are also based on statistical translation methods: the RECalign technique, which is based on the concept of recursive alignment, and the GIATIalign technique, which is based on simple word alignments. Experimental results are obtained for the EuTrans-I English-Spanish task, in order to create new, shorter bilingual segments to be included in a translation memory database. The evaluation of these three methods has been performed comparing the bilingual segmentations obtained by these techniques with respect to a manually segmented bilingual test corpus. These results show us that the new method proposed here outperforms in all cases the two already proposed bilingual segmentation techniques.

This work focuses on a hybrid machine translation system from Spanish into Catalan called SisHiTra. In particular, we focus on its word translation disambiguation module, which has to decide on the correct translation of each ambiguous input word in accordance with its context. We propose the use of statistical pattern recognition techniques for this task and, in particular, multinomial Naive Bayes text classifiers. Extensive empirical results on the use of these classifiers are presented, in which the influence of the window (context) size and parameter smoothing are carefully studied.

Bilingual documentation has become a common phenomenon in many official institutions and private companies. In this scenario, the categorization of bilingual text is a useful tool, that can be also applied in the machine translation field. To tackle this classification task, different approaches will be proposed. On the one hand, two finite-state transducer algorithms from the grammatical inference domain will be discussed. On the other hand, the well-known naive Bayes approximation will be presented along with a possible modelization based on n -gram language models. Experiments carried out on a bilingual corpus have demonstrated the adequacy of these methods and the relevance of a second information source in text classification, as supported by classification error rates. Relative reduction of 29% with respect to the best previous results on the monolingual version of the same task has been obtained.

We describe an algorithm to measure the similarity between sentences, integrating the edit distance between trees and single-term similarity techniques, and also allowing the pattern to be defined approximately, omitting some structural details. A technique of this kind is of interest in a variety of applications, such as information extraction/retrieval or question answering, where error-tolerant recognition allows incomplete sentences to be integrated in the computation process.

The automatic classification of music files into styles is one challenging problem in music information retrieval and for music style perception understanding. It has a number of applications, like the indexation and exploration of musical databases. Some techniques used in text classification can be applied to this problem. The key point is to establish a music equivalent to the words in texts. A number of works use the combination of intervals and duration ratios for music description. In this paper, different statistical text recognition algorithms are applied to style recognition using this kind of melody representation, exploring their performance for different word sizes and statistical models.

A foot database comprising 3D foot shapes and footwear fitting reports of more than 300 participants is presented. It was primarily acquired to study footwear fitting, though it can also be used to analyse anatomical features of the foot. In fact, we present a technique for automatic detection of several foot anatomical landmarks, together with some empirical results.