Détection de mots clés dans un flux de parole

The automatic speech recognition currently arouses a great interest. In particular, the keyword detection which constitutes a signi cant branch of the human-machine interaction and which can help us to communicate with our machines in a natural and direct way by using spontaneous speech. This technique consists in detecting in a pronounced sentence, the keywords characterizing the application and in rejecting out-of-vocabulary words as well as hesitations, false starts etc. The work presented in this thesis report deals with problem of keyword detection in a speech ow. First of all, we propose new garbage models founded on the modelling of the out-ofvocabulary words. Second, we introduce the recognition based on loop of phonemes, in which we apply various reward functions supporting keyword recognition. Then we propose to use con dence measures in order to make the decision of rejection or acceptance of a given keyword. The various con dence measures used are based on the probability of the local acoustic observation. First, we use these probabilities to calculate the arithmetic, geometric and harmonic means as con dence measures for each keyword. Second, we propose some others con dence measures based on the loop of phonemes recognition method. Finally we present the problem of detection as a classi cation problem where each keyword can belong to two di erent classes, namely correct and incorrect . This classi cation is carried out by using Support Vector Machines (SVM) which constitute a new technique of statistical training. Each recognized keyword is represented by a characteristic vector which constitutes the entry of the SVM classi er. To determine this vector, we use the probability of the local acoustic observation and then we introduce the duration of each state. In order to improve performances, we propose hybrid approaches combining the garbage models with the con dence measures and the con dence measures with the SVM. To test our models we use the French data base SPEECHDAT. The evaluation of all results was carried out using the ROC and the recall/precision curves. The best results were obtained by the SVM based methods. The hybrid methods also enabled us to get good performances.