| 摘要 |
<p>PURPOSE:To accurately determine the area of a registered person from only the learnt data of the registered person by using a neural network which constitutes an identification space according to the distance to input data. CONSTITUTION:A speaker is collated by using the speaker collation device equipped with a voice input part 11, a preprocessing part 12, the neural network, and a decision part 14 and the neural network 13 which constitutes the identification space according to the distance to the input data is used. Input voices 'tadaima(I'm home!)' of, for example, five males are divided equally into four on the time base and processed by Fourier analysis using a specific sampling frequency, frame length, and a frame period respectively, and power vectors of 16ch found in a specific band are averaged in respective blocks to obtain 64-dimensional input vectors. Those vectors are inputted to the neural network 13 and learnt sufficiently by respective units of an output layer corresponding to the speakers. Therefore, only the learnt data of the registered person are only learnt.</p> |
