Aggregating discriminative embedding by triple-domain feature joint learning with bidirectional sampling for speaker verification. (May 2023)

• Record Type: Journal Article
• Authors: Zi, Yunfei; Xiong, Shengwu
• Is Part Of: Biomedical signal processing and control. Volume 83(2023)
• Publication Date: 2023-05
• Subjects: Discriminative embedding -- Triple-domain feature -- Joint learning -- Bidirectional sampling -- Speaker verification; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2023.104703
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Improving speaker recognition performance. Proposing a triple-domain feature joint learning to aggregate the discriminative embedding for speaker verification. Proposing bidirectional sampling multi-scale network for capturing the different stages, different scale effective information. Design the Fisher feature fusion method, further enhance speaker individual information and reduce speaker commonality information. Designing the TribiNet model can achieve a robust model, high performance, and accuracy for speaker recognition. Abstract: Each axis of the speech, including time-domain, frequency-domain, and spectral-domain data, that's represents different physical meanings and different dimension information. Time-domain focus on physical signal versus time, the frequency-domain focus on the amount of signal in a given frequency band, and the spectral-domain focus on global power with speech, so only using the spectrogram to represent the whole information of speech to do speaker recognition will lose a lot of details information with the other dimensions. To tackle this limitation, we propose a triple-domain feature joint learning to enhance discriminative embedding from more dimensions for text-independent speaker verification. To further aggregate discriminative embedding, each domain uses a novel bidirectional sampling multi-scale feature aggregation network based on Fisher feature fusion to project spectrum features to more discriminative embeddings, termed TribiNet. Extensive experiments are conducted on a text-independent speaker verification dataset generated from the VoxCeleb corpus. The results demonstrate that the proposed method outperforms the state-of-the-art deep embedding architectures by at least 12%-58% in the test set. The ablation experiments further illustrate that our proposed approaches can achieve substantial improvement over prior methods.