HYPOTHETICAL RESCORING APPARATUS WHICH PERFORMS HYPOTHETICAL RESCORING IN SPEECH RECOGNITION SYSTEM

摘要

PROBLEM TO BE SOLVED: To rescore a hypothesis of an ASR (Automatic Speech Recognition) system by efficiently reflecting a wide speech context while lowering a data amount necessary for training of a speech model.;SOLUTION: A hypothetical rescoring apparatus includes a memory storing first, second and third acoustic models. The acoustic model corresponds to probability density functions p(X|A), p(X|A^-) and p(X|A⁺) of an observed value X when a central phoneme basic unit A, a multi-phoneme context A^- preceding the central phoneme basic unit A and a multi-phoneme context A⁺ following the central phoneme basic unit A are given. The apparatus further includes a rescoring unit calculating a given hypothetical acoustic score by an expression (1). The acoustic score of the j-th phoneme is given by an expression (2). A_j is the central phoneme basic unit including a given hypothetical j-th phoneme, X_j is a j-th observed value unit separated in a given hypothesis, and N is the number of the phonemes in the hypothesis.;COPYRIGHT: (C)2007,JPO&INPIT