Purpose There are upper airway abnormalities in patients with obstructive sleep apnea (OSA), and their speech signal characteristics are different from those of unaffected people. In this study, the severity of OSA was evaluated automatically by machine learning technology based on the speech signals of Chinese people. Methods In total, 151 adult male Mandarin native speakers who had suspected OSA completed polysomnography to assess the severity of the disease. Chinese vowels and nasal sounds were recorded in sitting and supine positions, and the accuracy of predicting the apnea-hypopnea index (AHI) of the participants using a machine learning method was analyzed based on features extracted from the speech signals. Results Among the 151 participants, 75 had AHI > 30 events/h, and 76 had AHI <= 30 events/h. Various features including linear prediction cepstral coefficients (LPCC) were extracted from the data collected from participants recorded in the sitting and supine positions and by using a linear support vector machine (SVM); we classified the participants with thresholds of AHI = 30 and AHI = 10 events/h. The accuracies of the classifications were both 78.8%, the sensitivities were 77.3% and 79.1%, and the specificities were 80.3% and 78.0%, respectively. Conclusion This study constructed a severity evaluation model of OSA based on speech signal processing and machine learning, which can be used as an effective method to screen patients with OSA. In addition, it was found that Chinese pronunciation can be used as an effective feature to predict OSA.