Endangered Tujia Language Speech Enhancement Research Based on Improved DCGAN

Abstract

As an endangered language, Tujia language only rely on oral communication. There must exist noises in the process of collecting Tujia language corpus. This paper studies an end-to-end speech enhancement model based on improved deep convolutional generative adversarial network (DCGAN) to extract nearly pure Tujia language speech in noisy environment. Due to the low resource nature of Tujia language, using Chinese corpus as an extension of the Tujia language can effectively solve the problem of insufficient data. The speech enhancement function of the Tujia language was realized using the end-to-end method that consists of symmetric encoding and decoding. By modifying the loss function and network hierarchy parameters, adding the spectrum normalization and imbalanced learning rate made the model more stable during the training process. The experimental results show that the speech enhancement method proposed in this paper can achieve better noise reduction effect on the Tujia language dataset than traditional speech enhancement algorithm and neural network enhancement algorithms.