By means of a small number of conductive patterns adaptive reference pattern adaptation in front of the direction of

摘要

PURPOSE:To obtain a high-accuracy speaker adaption system which does not require a large amount of voicing data of many speakers by correcting feature parameters featuring standard patterns of respective categories by using learning data having plural elements corresponding to them. CONSTITUTION:Voicing data (standard data) of a standard speaker which is used to generate a hidden Markov model(HMM) before adaption and voicing data (learning data) of an unknown speaker who voice the same contents are prepared and matched and a data correspondence process for making respective elements to correspond one to one is performed. Then the standard data is segmented and which of states of the HMM the elements of the standard data are in is classified. Then, the similarity between the mean vectors of the respective states and the element of the standard data is found. After the preparations, difference vectors between the mean vectors in the respective states after and before the adaption are found and used to find a mean vector after the adaption which is a data-presence category correction feature parameter.