| 摘要 |
PURPOSE: A method for expressing Gaussian probability density and a voice recognition training method are provided to easily obtain HPGAM(Hybrid Partitioned Gaussian Autoregressive mixture) having a better recognition ratio than a prior GAM(Gaussian Autoregressive mixture) and PGAM(Partitioned Gaussian Autoregressive mixture). CONSTITUTION: A probabiity density space is divided (401) into many PGAM spaces. The divided PGAM space is expressed (402) as a GAM. The highest GAM is expressed (403) as HPGAM, thereby expressing a Gaussian probability density. A voice recognition model in which the nuimber of mixture groups are initialized, a trained voice recognition model is obtained by a recognition training. A voice recognition test is made to the trained voice recognition model, a recognition tranining is made about a mis-recognition word. A new voice recognition model is obtained. A new voice recognition model having the increased number of mixture groups is obtained. A recognition training is performed to the obtained voice recognition model, and it is checked that the number of groups reaches to a desired number. If the number of groups reaches to the desired number, a voice recognition training is terminated. |
