The speech recognition feature extraction methods used nowadays are not optimal when they are applied to specific environment and specific recognition task. To deal with this problem, the new concepts of regional characteristic and trace characteristic are proposed, accompanied by the definition of the new parameter of regional resolution in the speech feature vector space. Under these conceptions, optimization is introduced to turn the previous non-optimal, universal feature extraction method into the new optimal, task-dependent and environment-dependent one, which will improve speech recognition result without changing substructure of the original method or increasing the computation complexity.
展开▼
