Improved detection of boundaries of phonemes in speech databases

An algorithm for improvement of boundaries estimated via HMM is suggested, using numerous divergence measures (spectral distance, cepstral distance, difference of the cepstral coefficients, Kullback-Leibler divergence, distance given by the General Likelihood Ratio, distance defined by the family of Recursive Bayesian Changepoint Detectors, the Bhattacharyya divergence, the Mahalanobis measure, the L2 metric and the Jeffreys-Matusita measure). A method for choosing of appropriate measure for each boundary is stated, and is demonstrated on the training set of signals. The ability to improve boundary positions is verified on the testing set, and it is found that a combination of methods brings better results than using only one measure.