Aiming for best fit t-norms in speech recognition

Here we generalize the model of automatic speech recognition (ASR) based on the maximization of products of probability likelihoods of each corresponding speech frame and phoneme by applying strict t-norms. We formulate it as a minimization problem in terms of the logarithmic generator of strict t-norms and investigate the experimental solutions for piecewise linear logarithmic generators. The performance of the best fit t-norms found in this manner for a database used earlier proved to be superior than that of classical t-norms.