Extraction of Noise-Robust Speaker Embedding Based on Generative Adversarial Networks
Jianfeng Zhou, Tao Jiang, Qingyang Hong, Lin Li · 2019
In the field of speaker verification, the speaker systems based on x-vector framework are widely used in many scenarios. However, it suffers from the performance degradation caused by noise disturbance. In this paper, we firstly analyzed the noisy robustness of x-vector by training the networks using a mixture dataset which includes clean data and corrupted data. Then, we proposed a novel adversarial strategy against noise interference and extracted the noise-robust speaker embedding with x-vector. The proposed adversarial method named as triple-net GAN employs three connected networks: a generator network (G), a discriminator network (D) and a classifier network (C). The spectral coefficients of clean and noisy speech utterances are fed to the G, of which the structure is nearly the same as x-vector. The outputs of G are transferred in a parallel way to the D and C. And the labels of D are set binary for clean data and corrupted data, while the labels of C are set corresponding to speaker identities, which aims to learn the speaker embedding features invariant to the noise. Finally, we executed the experiments with different variants of triple-net GAN to verify the denoising capability of the proposed adversarial method. Experimental results on Librispeech corpus demonstrate that our proposed method could achieve a better performance under the noisy environments.