Multi-environment characteristic compensation method for voice recognition system
暂无分享,去创建一个
The invention discloses a multi-environment characteristic compensation method for a voice recognition system. In a training phase, firstly, multi-environment voice is acquired to obtain training voice of a plurality of fundamental training environments, and secondly, a voice model for generating each fundamental training environment is trained by the aid of the training voice of the fundamental training environment. In a recognition phase, firstly, the voice model of the fundamental training environment the closest to a testing environment is selected for current testing voice, secondly, parameters of the selected voice model are transformed so that the selected voice model is matched with the current testing environment, and finally, the transformed voice model is used for estimating pure voice characteristic vector from the testing voice with noise. By the aid of the multi-environment characteristic compensation method, the performance of the voice recognition system in severe application environments with low signal-to-noise ratio and the like can be remarkably improved, and robustness of the system is enhanced.