Application to a noise reduction filter for hearing aid of cascaded sandglass-type neural networks

摘要

A noise reduction filter for speech was constructed of cascaded sandglass-type neural networks (CSNNRF).The number of cascade was adaptively determined by estimated power of noise contained in an input signal. It was shown by bearing tests that the CSNNRF improved the intelligibility of speech signals however power of noise varied. Dynamics of CSNNRF was examined when speech signal was processed. When a vowel was put into a CSNNRF, it gained dynamics that passed the low frequency range including the lowest formant of the vowel. When a consonant was put into a CSNNRF, the dynamics was remarkably affected by a preceding vowel. It bad a low gain in the high frequency range to distort the consonant. When a S/N ration of input was high, CSNNRF got dynamics with a low gain in the high frequency range after a consonant input, while it got dynamics with a flat gain when a S/N ratio was low.