Abstract
The purpose of this research is to improve the performance of our lab’s embedded voice command system. The goal is to speed up the processing time and reduce the additional errors caused by our method. In order to do so, we propose two methods to reduce the feature dimension required by the system as follows. The first one is to directly reduce original 39 dimensions. The second one is to use heteroscedastic linear discriminant analysis after increasing the dimension of original feature vectors. Then, we change the floating-point transform matrix to a fixed-point version through a scale factor and store it in the system for the feature transformation in runtime. Based on the second method, different parameter settings are tested. The final experimental result shows that the second method (heteroscedastic linear discriminant analysis) outperforms the first method (direct feature reduction). The first method even performs better than the original method with 39 dimension of feature in some cases. This result indicates that heteroscedastic linear discriminant analysis is able to effectively accelerate the recognition time while at the same time reduce the error rate. However, the experimental results also show that the performance changes with different number of mixture components in acoustic models and the analysis method. We can therefore choose the most suitable way to do the analysis for the best performance.