| 摘要 |
Hitherto, when speech data is quantized, since a quantum is determined irrespective of a deviation of a distribution of the speech data, an error of an output probability calculation by quantization is large and a recognition rate deteriorates. According to the invention, to solve the subject of the prior art, a quantization range of speech data is obtained, an integral value of an output probability of the quantization range is obtained by using an output probability distribution of statistic models of speech data which has previously been obtained, a quantum is determined so as to equally divide the integral value, and a quantization code book is formed, thereby reducing an error of an output probability due to the quantization and enabling a recognition rate to be improved.
|
