A construction of noise reduction filter by neural network - application for digital hearing aid appliance as preprocessing

摘要

We proposed an adaptive noise reduction filter composed of Cascaded Sandglass-type Neural Network (CSNNRF) to develop a hearing aid appliance. It was known that Sandglass-type Neural Network (SNN) has property of processing signals which is equivalent to KL transformation - KL inverse transformation after learning. First, we calculated the expected value of the speech element from KL transformation element of input signal. We decided coefficients in proportional to the expected value and multiplied them to the output of each stage of CSNNRF. Next, the sound volume was adjusted by multiplying amplification rate to the total sum of output signals of CSNNRF. Here, amplification rate was determined as the inverse of the sound volume. Finally in order to improve S/N ratio, the number of stages in CSNNRF was altered adaptively. As a result of the noise reduction experiment, it was shown that the noise was reduced adaptively and the sound volume was adjusted appropriately.