To apply the visual image analysis of speech spectrogram to human fatigue detection effectively, a human fatigue detection method based on speech spectrogram features is proposed. Firstly, the influence mechanism analysis of human fatigue on speech spectrogram is analyzed. The Mel frequency stretching transform of speech spectrogram based on the auditory perception theory is used to highlight the region of interest which is susceptible to fatigue. Secondly, the Mel frequency stretched spectrogram is divided into 24 overlapping critical frequency band sub-images, and 15 texture features are extracted from the gray level co-occurrence matrixes of each sub-image in 4 directions to quantitatively describe the fatigue information. Finally, a human fatigue detection model based on multi sub-bands fatigue information fusion is formulated by designing the feature-layer classifier for distribution detecting the features of each critical frequency band. In this way, the fatigue detection result can be achieved, which is based on the decision-level multi-classifiers fusion decision. Experimental results show that the extracted speech spectrogram features have stronger fatigue classification ability than traditional acoustic features. The fatigue detection effectiveness of this method is also better than the existing spectrogram feature recognition methods.