Improving timbre recognition and Internet technology have created new avenues for developing vocal teaching models in colleges and universities. In this paper, several people propose a timbre-based Internet vocal teaching model based on the cross-classification vocal teaching model. The timbre is extracted and divided into training and test data, dimensionality reduction and classification processes to form the general process of timbre recognition. A GMM-HMM timbre recognition model is proposed by describing the observed timbre probabilities using GMM so that the observed timbre states of HMM are continuous. The SVM classification algorithm is employed to categorize the identified timbres and determine the specific distribution of incorrect timbres. In the vocal teaching test, the Mayer inversion coefficient achieved a higher RPA when the number of candidate pitches was less than 4. If the number of candidate pitches exceeds 5, the amplitude-free weighting method will increase slowly.