NOISE ADAPTATION SYSTEM OF VOICE MODEL, NOISE ADAPTATION METHOD, AND NOISE ADAPTATION PROGRAM OF VOICE RECOGNITION

摘要

PROBLEM TO BE SOLVED: To simplifying voice recognition processing and reducing a calculation amount.;SOLUTION: Models of each noise are learned by using a noise database. Distances among each noise model are calculated and clustering of the noises is performed. Next, voice models is created in terms of a tree structure according to a result of the clustering of the noises. At first, the noises are clustered by SNR (signal to noise ratio), next, tree structure models are prepared at each SNR condition, and space of noise/voice models in the tree structure are stored into a tree structure model storage part 1. A feature parameter sequence is extracted by analyzing input noise voices to be recognized in a feature extracting process by a feature extracting part 2, and an optimum model is selected by a model selection determining part 4 from the space of the noise/voice models. An adaption part 5 of a model linear transformation performs linear transformations so that likelihood is further maximized from the selected space of the noise/voice models.;COPYRIGHT: (C)2004,JPO