摘要 |
PURPOSE:To prevent the occurrence of degradation in precision caused by a bit omission and a quantization error by using learning functions of neural networks. CONSTITUTION:When voice signals are inputted from a voice input section 5, a linear prediction analysis section 15 computes linear prediction coefficients for the amount of analysis dimensions from the inputted voices which are sampled at a specified time interval, a linear prediction coefficient quantization section 16 quantizes these linear prediction coefficients, a linear predictor 17 computes linearly predicted voices based on the quantized signals and the information from a code book. A hearing sensation weighted filter 18 reduces noises against linearly predicted errors which are difference values between the inputted voices and linearly predicted voices, an average square error computing section 19 computes average square errors and holds a minimum average square error and the stimulate vector at that time. And a zero condition response computing section 13 receives the stimulate vector and linear prediction coefficients, computes the response value only due to the stimulate vector and outputs the difference values between the input voices and the response values as the teacher's data of two layers layer type neural network. |