Modèles de langue exploitant la similarité structurelle entre séquences pour la reconnaissance de la parole
Christian Gillot · HAL (Le Centre pour la Communication Scientifique Directe) · 2012
The role of a stochastic language model is to give the best estimation possible of the probability of the sequence of words in a given language. It is an essential component of any speech recognition software and has a great influence on performance. The state-of-the-art models most commonly used are the n-gram models smoothed using the Kneser-Ney technique. These models use occurrence statistics of word sequences typically up to a length of 5, statistics computed on a large training corpus. This thesis starts by an empirical study of the errors of a state-of-the-art speech recognition system in French, which shows that there are many regular language phenomena that are out of reach of the n-gram models. This thesis thus explores a dual approach of the prevailing statistical paradigm by using memory models which process efficiently specific phenomena, in synergy with the n-gram models which efficiently main trends. The notion of similarity between long n-gram is studied in order to identify the relevant contexts to take into account in a first similarity language model. The data extracted out of the corpus is combined via a Gaussian kernel to compute a new score. The integration of this non-probabilistic model improves the performance of a recognition system. A second model is then introduced, probabilistic and thus allowing a better integration of the similarity approach with the existing models and improves the performance in perplexity on text.