To prevent degradation of the sound quality by suppressing a noise in an input signal as conventional and limiting suppression to mitigate the distortion with respect to a voice or a tone.

A first state decision part 11 decides the state of an input sound signal to output a state decision signal corresponding to the decision result. A parameter calculation part 12 calculates a linear prediction coefficient and variance as model parameters of sounds and noise, which are used in a Kalman filter 13, and determines values of these linear prediction coefficient and variance as prescribed values in accordance with the state decision signal. A second state decision part 14 changes the state decision signal from the first state decision part 11 in accordance with the magnitude of the residual energy of the input signal obtained from the result of linear prediction analysis of the parameter calculation part 12. The parameter calculation part 12 changes determined values of the linear prediction coefficient and variance from prescribed values to another prescribed values in accordance with the changed state decision signal and outputs them.