The embodiment of the invention discloses a speech recognition method and device. The method specifically comprises the steps that on the basis of acoustic scores of speech frames in speech signals, non-silent frames in the speech signals are determined; buffer frames are determined between the adjacent non-silent frames based on the acoustic scores of the voice frames, wherein a modelling unit corresponding to the buffer frames represents the beginning or end of a statement; on the basis of speech frames obtained after the buffer frames are removed from the speech signals, decoding is conducted, and the speech recognition results are obtained. Accordingly, the number of the speech frames needing to be decoded can be effectively reduced, the operation pressure of a decoder is lowered, thedecoding speed is increased, and then the speech recognition speed is increased.