Recognizing teaching roles in classrooms is a primary work in teaching analysis. Previous work utilized speaker diarization systems to distinguish “who is speaking when” in a scene. However, in complex classroom scenes, the effect of the proposed method may be limited by an unknown number of speakers, microphone interference, etc. To construct a teaching role recognition method suitable for classroom scenes, this paper found a commonality of speakers with the same teaching identity at the acoustic level. Based on this commonality, this paper adjusted the speaker recognition task and applied it to recognize the teaching identities of different speakers in classroom scenes. A neural network named SA-TDNN was constructed to embed speakers. SA-TDNN achieved 10.09% on CN-Celeb, which is more accurate than other methods. SA-TDNN achieved 13.25% on the self-built dataset, which is more accurate than other methods. In addition, the proposed method outperforms the existing general SD system by 3.83% on the validation set of self-built dataset, which supports the feasibility of this method in the Chinese classroom.