부산대학교 도서관

Self-supervised automatic speech recognition (ASR) is a technique that has gained popularity in recent years due to its ability to improve ASR accuracy using large amounts of unlabeled data. Language identification (LID) is another important technique that has been used to identify the language of speech segments, which can be useful for developing ASR models for multilingual environments. This paper presents a novel approach to self-supervised ASR using LID, where the ASR model is pre-trained on large amounts of unlabeled data, and the language of the speech segments is identified during training. The identified language is then used to guide the training of the ASR model, resulting in improved performance on both in-domain and out-of-domain data. The proposed approach is evaluated on several benchmark datasets, demonstrating its effectiveness in improving ASR accuracy and reducing language mismatch. The results suggest that self-supervised ASR using LID has great potential for developing more robust and accurate ASR systems in multilingual environments.