This article is based on the application research of multimedia feature extraction technology in the development of vocal performance teaching in universities. Combining with the core image and sound modules in feature extraction technology, this article proposes an application model of multimedia feature extraction technology based on image HOG algorithm and Mel spectrum for vocal audio recognition in vocal performance teaching in universities. Experiments have shown that the features extracted by this method can not only effectively identify the styles of different types of singing works, but also recognize the personality characteristics of singers. At the same time, it can effectively reduce the misclassification rate caused by noise interference, thereby improving the recognition rate.