Exploiting Adaptive Adversarial Transfer Network for Cross Domain Teacher's Speech Emotion Recognition

The speech of teachers in classroom teaching contains their teaching emotions, which have a direct impact on the quality of classroom teaching and the learning effectiveness of students. Therefore, emotional recognition of teacher’s speech will help improve their self-awareness and provide scientific support for improving teaching methods. However, in real-world scenarios, many data are unlabeled, especially in the emerging field of smart education, which lacks high-quality teacher speech datasets. Thus, a cross domain teacher’ speech emotion recognition model based on adaptive adversarial transfer networks is proposed, which dynamically evaluates the relative importance of global and local distributions using adaptive adversarial modules and automatically assigns weights to both. The common features of the source and target domains were obtained to complete cross domain speech emotion recognition for teacher classroom teaching. Finally, the experiment showed that the classification accuracy of the proposed model reached 75.3%, significantly higher than other cross domain recognition models.