The disclosure discloses a speech recognition method and a device, aiming at recognizing words in a sentence text after determining the sentence text corresponding to an input speech, and substituting after determining the wrong words which do not conform to an application scenario in the sentence text, so as to improve the accuracy of speech recognition. The speech recognition method according to embodiments of the present disclosure includes: recognizing a sentence text corresponding to an input speech according to the speech; recognizing wrong words in the sentence text; determining substitute words corresponding to the wrong words; and substituting the wrong words with the substitute words. The determination of the wrong words is based on comparing an average similarity of the word with N other similar words against a first threshold and comparing a maximum similarity between the said word and other words in the sentence against a second threshold.