Voice recognition method and system

The embodiment of the invention provides a voice recognition method. The voice recognition method comprises the steps that audio features of each frame of an extracted voice file are input into a deeplearning neutral network, the posterior probability of each frame is determined, and by smoothing the posterior probability of each frame, key words constituting conversation voice are determined; astring word set where the key words are located is determined; a first label sequence composed of labels corresponding to the maximum posterior probabilities of all the frames in the voice file and asecond label sequence determined by pronunciation mapping of all to-be-selected words are obtained, similarities between the first label sequence and the second label sequence corresponding to all theto-be-selected words are traversed, and the to-be-selected words corresponding to the maximum similarity serve as recognized words of the conversation voice. The embodiment of the invention further provides a voice recognition