Speech recognition device, acoustic model learning device, speech recognition method and acoustic model learning method

摘要

A speech recognition device comprises: a feature-space transformation matrix estimation unit (6) for estimating a plurality of feature-space transformation matrices corresponding to a regression tree via linear regression for feature space using a posterior probability for a first recognition result acquired by speech recognition, a regression tree created from an acoustic model, and a speech feature extracted from an input speech; and a feature conversion unit (7) for converting a speech feature using a feature-space transformation matrix determined on the basis of the acoustic model and the state series in the acoustic model from among the plurality of feature-space transformation matrices. A decoding unit (2) compares the converted speech feature with the acoustic model to perform speech recognition and thereby obtain a recognition result.