For speech detection in Parkinson’s patients, we proposed a method based on time-frequency domain gradient statistics to analyze speech disorders of Parkinson’s patients. In this method, speech signal was first converted to time-frequency domain (time-frequency representation). In the process, the speech signal was divided into frames. Through calculation, each frame was Fourier transformed to obtain the energy spectrum, which was mapped to the image space for visualization. Secondly, deviations values of each energy data on time axis and frequency axis was counted. According to deviations values, the gradient statistical features were used to show the abrupt changes of energy value in different time-domains and frequency-domains. Finally, KNN classifier was applied to classify the extracted gradient statistical features. In this paper, experiments on different speech datasets of Parkinson’s patients showed that the gradient statistical features extracted in this paper had stronger clustering in classification. Compared with the classification results based on traditional features and deep learning features, the gradient statistical features extracted in this paper were better in classification accuracy, specificity and sensitivity. The experimental results show that the gradient statistical features proposed in this paper are feasible in speech classification diagnosis of Parkinson’s patients.