The invention discloses a vocal cord anomaly detection method based on acoustic phonetic features. The method comprises the steps of firstly, extracting mel frequency cepstral coefficient (MFCC), the fundamental frequency F0, the fundamental frequency perturbation Jitter, the amplitude perturbation Shimmer and the harmonics to noise ratio HNR for each frame of voices; adopting acoustic features as an input, and respectively training a Gaussian mixture model theta A and a Gaussian mixture model theta N based on the expectation-maximization EM algorithm, wherein the Gaussian mixture model theta A and the Gaussian mixture model theta N respectively represents the vocal cord abnormal state and the vocal cord normal state; finally, respectively inputting the feature matrix F of the test voice into the Gaussian mixture model theta A and the Gaussian mixture model theta N so as to obtain a corresponding output probability P (F|theta A) and a corresponding output probability P (F|theta N). If the P (F|theta A) is larger than the P (F|theta N), the vocal cord of the speaker of the test voice is abnormal. Otherwise, the vocal cord of the speaker of the test voice is normal. According to the technical scheme of the invention, multiple sets of acoustic features, adopted as the inputs of Gaussian mixture models, are extracted from the test voice, wherein the acoustic features can effectively reflect the state of the vocal cord. Therefore, voices in the vocal cord abnormal state and in the vocal cord normal state can be effectively distinguished. As a result, whether the vocal cord of the speaker of the test voice is abnormal or not can be judged. The method has the advantages of non-intrusion, convenience, low cost and the like.