Speech model learning method, noise suppression method, speech model learning apparatus, noise suppression apparatus, speech model learning program, and noise suppression program

摘要

PROBLEM TO BE SOLVED: To improve noise suppression performance.SOLUTION: A noise suppression system 200 records a mixed acoustic signal having an audio signal and noise signal mixed, and extracts a feature quantity, which is used to perform noise suppression, and a feature quantity, which is used to calculate a speech posteriori probability, from the mixed acoustic signal. The noise suppression system 200 uses a logarithmic mel spectrum, speech GMM, normalized logarithmic mel spectrum, and speech DNN resulting from learning to estimate a speaker adaptation parameter and a parameter set of a noise GMM that is a probability model of noise. The noise suppression system 200 uses a complex spectrum, logarithmic mel spectrum, speech GMM, speaker adaptation parameter, parameter set of the noise GMM, and speech posteriori probability to construct a noise suppression filter, and suppresses noise so as to obtain a noise suppressed signal.SELECTED DRAWING: Figure 4