Denoising automatic encoder training method and speaker recognition system

Abstract

The invention discloses a denoising automatic encoder training method, which adopts an automatic encoder comprising six full connection layers. The denoising automatic encoder training method comprises the steps of: carrying out noise adding processing on sample voice data to obtain hybrid voice data; performing embedded feature extraction on the hybrid voice data to obtain hybrid embedded features; performing embedded feature extraction on the sample voice data to obtain sample embedded features; inputting the hybrid embedded feature into the denoising automatic encoder to obtain denoising embedded features; and training the de-noising automatic encoder by minimizing the difference between the de-noising embedded features and the sample embedding. According to the denoising automatic encoder training method of the invention, the denoising auto-encoder is trained, the noise-added user voice features are used as input, and the clean user features are used as tags for training, so that the auto-encoder is used for denoising. The robust performance to noise can be improved, and the accuracy rateof identifying users in a noisy environment is improved.

Description

technical field [0001] The invention relates to the technical field of artificial intelligence, in particular to a denoising automatic encoder training method and a speaker recognition system. Background technique [0002] Much progress has been made in speaker verification (SV) with the development of benchmark datasets and speaker embeddings extracted by deep neural network (DNN) training. For example, DNN-based embeddings of d-vectors and x-vectors have outperformed traditional i-vector systems trained on generative shallow models. In particular, the utilization of large-scale datasets and data augmentation lead to state-of-the-art performance of x-vectors on the SV task. Although the x-vector system achieves excellent performance on relatively clean datasets (e.g., VoxCeleb and SRE), significant performance degradation can still be observed for real applications with complex noise sources. Therefore, building a SV system with strong noise immunity is still a daunting t...

AI Technical Summary

Problems solved by technology

Although the x-vector system achieves excellent performance on relatively clean datasets (e.g., VoxCeleb and SRE), significant performance degradation can still be observed for real applications with complex noise sources

Therefore, building a SV system with strong noise immunity is still a daunting task

Images