SPEECH RECOGNITION DEVICE, SPEECH RECOGNITION METHOD, AND SPEECH RECOGNITION PROGRAM

摘要

The present invention can increase the types of noise that can be supported by speech recognition having a highly accurate speech recognition rate. A speech recognition device: stores suppression coefficients in association with adaptation coefficients, each said suppression coefficient indicating the amount of noise suppressed, each said adaptation coefficient indicating the amount of application of a noise model, which is created on the basis of predetermined noise, to be combined with a clean sound model, which is created on the basis of speech that does not include any noise; estimates noise from an input signal; removes from the input signal an amount of noise, out of said estimated noise, that is specified by the suppression amount determined on the basis of the suppression coefficient; combines a noise model created on the basis of the estimated noise with said clean sound model in accordance with the application amount determined on the basis of the adaptation coefficient, and thereby creates a post-adaptation sound model to which noise has been applied; and recognizes speech on the basis of the noise-suppressed input signal and the created post-adaptation sound model.