Symmetric structure for teacher-student home acoustic scene domain adaptation

This paper mainly addresses the problem of acoustic domain adaptation of speech recognition models in smart home scenarios, and for the interference problem of home noise on speech recognition, this paper proposes a symmetric structure of Teacher-Student training paradigm, specifically by allowing the output of the Student model trained on the background noise dataset of the quiet audio mixing to be close to the output of the quiet audio dataset trained on the Teacher model, thus achieving the purpose of domain adaptation for home scenarios. The form of Teacher model's output implicitly allows the noise scene recognition results to converge to the recognition results of the quiet scene, so as to achieve the purpose of domain adaptation to the home scene, so that it can show higher accuracy in the home acoustic environment. In this paper, we use the Mandarin speech dataset mixed with the noise dataset to construct the background noise data of the home scene, and the comparison experiments in the noisy environment show that the word error rate of the Conformer model obtained from the training of the proposed Teacher-Student model decreases by 6.21% compared with that of the Conformer baseline model in the noisy environment.