A Intelligent Speech Recognition Method Based on Stable Learning

Abstract Speech is the main way of human communication, which carries both the speaker's information and the speaker's emotion. A variety of applications can harness emotion in speech to serve human needs more effectively. The deep learning algorithm is a practical solution to the classification nature of speech recognition. Various algorithms have been widely utilized for voice data and achieved remarkable performance. However, in real life, the testing data to be under a different distribution from the training data, this will cause the out-of-distribution(OOD) problem. This article proposes a new domain generalization method for speech classification based on Stable Learning (StableNet) to address the OOD problem. The StableNet can remove the connection between features through learning weights for training samples, which makes deep models learn more useful features instead of the fake connection between the discriminative features and labels. We evaluate the performance of the proposed method by conducting speech classification experiments on voice datasets. We also investigate the importance of various features on speech classification in noisy environments. The effects of proposed method on speech recognition performance are evaluated.