The invention relates to the field of speech recognition, in particular to a speech recognition method and a video conference system based on speech recognition. The method comprises the following steps that speech signal samples are denoised through a wavelet denoising threshold function; speech signal samples obtained after denoising are pre-processed; the pre-processed speech signal samples aresubjected to feature extraction, and a model is obtained through training; to-be-recognized speech signals are subjected to speech matching through the model obtained through training, and recognizedtext is output. By means of the established wavelet denoising threshold function, not only can more useful signals be retained, but also processing can be better conducted in smoothness, so that a better denoising effect can be achieved, and the accuracy of speech recognition is further improved.