Modelamento de coeficientes de adaptação pra sistemas de reconhecimento automátifco de fala

Abstract

The mismatch between the acoustic conditions of the training utterances and those experienced by automatic speech recognition systems is one of the responsible factors for its performance degradation when operating in noisy environments. This is a relevant issue in the current reality with increasing use of these systems on mobile devices. Among the various techniques proposed in the literature to minimize this challenge, the adaptation based on Maximum a Posteriori criteria (MAP) stands out where the acoustic models from training stage can be adapted to the noise condition (type and level) experienced by the system. In this approach, noise samples are used to modify the parameters of the acoustics models to maximize the word accuracy. The intensity of this modi¯cation depends on an adaptation coeficient which is usually calculated empirically through a grid search. In this dissertation, a modeling of how the great values of these coe±cients behave according the type and level of noise is performed. From this result, an algorithm to determine an appropriate value for it is proposed. It is based on the parametric adjustment by application of logistic curve minimizing the processing time. The adaptation coe±cient provided by this algorithm does not lead the maximum word accuracy for all cases, but it always provides gain. The experimental results show an gain of 3% on word accuracy.