| 摘要 |
<p>PURPOSE:To remove noises and to improve the S/N of a signal by using an analysis part which converts a signal containing noises into a feature vector and a neural network which outputs the index of a representative vector in a corresponding cluster when the feature vector is inputted. CONSTITUTION:The voice signal which contains noises is converted by the analysis part into an autocorrelation coefficient 02 and the maximum value 104 is detected as dummy power by a maximum value detection part 103 and normalized by a normalization part 105. The normalized autocorrelation coefficient 106 is inputted to the neural network 107 in hierarchic structure wherein the units of the input layer, intermediate layer, and output layer are weighted and coupled and the neural network 107 outputs the index 108 of the representative vector in the best corresponding cluster. The outputted index 108 is sent from a transmission part 109 to a reception part 111 through a transmission part 110 together with the maximum value 104 of the autocorrelation coefficient 102 detected by the maximum value detection part 103. Consequently, the S/N of the signal is improved.</p> |
