| 摘要 |
<p>The output probability of a mixture Gaussian HMM is given by a function, such as a mixture multi-dimensional Gaussian distribution. A mixture multi-dimensional Gaussian distribution is the sum of multi-dimensional Gaussian distributions, and a multi-dimensional Gaussian distribution is the product of one-dimensional Gaussian distributions of respective feature components. The values of distribution relating to the representative dispersions and averages of one-dimensional distributions are listed in a numeric value table (1052). For the feature component of a feature vector, which is an object to be recognized, linear scalar quantization is adopted, and by using the quantized value as index, information in intermediate tables (401, 402) is referred to. The intermediate tables are prepared by extraction from a global table (400). The global table contains X-direction arrays in the Y direction, each of which is an array of address information representing the locations of numeric values of each one-dimensional Gaussian distribution listed in the numeric value table. The value of the dispersion (ξ) is taken into consideration to select an X-direction array; the value of the average ((µ) is taken into consideration to select the first location in the X-direction of the array. In such a way, intermediate tables are extracted from a global table, and a numeric value table is referred to by using address information at the location where the quantized value is used as index according to the first location.</p> |
