A voice data enhancing method based on a recurrent neural network in the field of voice recognition processing aims at solving the problem of excessive modeling word dependence caused by irregular grammar phenomena of voice recognition simulation in voice recognition in a recurrent neural network. The method comprises: extracting acoustic features of various frequency energy values identifying voice from input voice data to generate acoustic feature vectors (201); obtaining a statement label sequence of the voice data according to a preset labeling file and the acoustic feature vectors (202); obtaining an alignment file after a decision cluster operation by means of the labeling file preset by a decision cluster, and the statement label sequence (203); generating a first random number γ between [0, 1], and comparing the first random number with a preset adjusting proportion α (204); and if the first random number γ is greater than the adjusting proportion α, performing enhancement processing on the voice data in a position indicated by a boundary file (205). The method enables irregular spoken language phenomena in training data to be increased quickly and conveniently.