Deep Speaker Verification Model for Low-Resource Languages and Vietnamese Dataset

Deep Speaker Verification Model for Low-Resource Languages and Vietnamese Dataset Vi Thanh Dat VNG corporation datvt7@vng.com.vn Pham Viet Thanh Hanoi University of Science and Technology thanh.pv.ds@gmail.com Nguyen Thi Thu Trang∗ Hanoi University of Science and Technology trangntt@soict.hust.edu.vn Abstract Speaker verification is an essential task in speech processing with great authentication and surveillance applications. Large-scale datasets have hugely contributed to the suc- cess of neural networks for speaker verifica- tion. However, in low-resource languages, building such massive datasets is infeasible. This paper aims at proposing a speaker verifi- cation model for low-resource scenarios, with the baseline from Clova’s H/ASP system in VoxSRC 2020. The proposed method adopts transfer learning to utilize the knowledge from the model trained with VoxCeleb, an English large-scale dataset. For network optimization, Stochastic Gradient Descent is employed in- stead of Adam because of its superior gen- eralization. This work also proposes a novel marginal variant of Angular Prototypical (AP) Loss, i.e. Angular Margin Prototypical (AMP) Loss, which encourages a more discrimina- tive embedding space. To experiment with the proposed model, we investigated building a public speaker verification dataset for Viet- namese. A processing pipeline is proposed to enhance the quality of the dataset. After being collected from sources, noisy speakers and noisy utterances are removed using self- similarity matrix analysis. Speakers with the same identity are then unified. The experi- mental results show that the proposed model achieves an Equal Error Rate (EER) of 3.1% which outperforms the baseline with 7.6% EER on the collected dataset. ∗*Corresponding author 1 Introduction Speaker verification is a task which takes an un- known speech as input and determines whether the speech matches the claimed identity. With the de- velopments of deep neural networks, speaker verifi- cation systems have gained huge advances and out- perform traditional probabilistic systems. The application of neural networks in speaker ver- ification is to learn discriminative speaker embed- ding space. Commonly used architectures in pio- neering works include TDNN-based models (Snyder et al., 2015) and networks architectures originated from computer vision researches such as VGG (Si- monyan & Zisserman, 2014) and ResNet (He et al., 2016). Since speaker verification is a metric learn- ing problem, the idea is to learn embedding features that have small intra-class distance and large inter- class distance. Several distinctive metric learning loss functions for speaker verification have been pro- posed, from the well-known Triplet Loss (Zhang et al., 2018) to Prototypical Loss (J. Wang et al., 2019) and Angular Prototypical Loss (Chung et al., 2020). The success of speaker verification systems does not come only from the advancement of neural net- works but is also attributed to the availability of large-scale speaker datasets. Good examples are VoxCeleb1 (Nagrani et al., 2017) and VoxCeleb2 (Chung et al., 2018) . These datasets together con- tribute more than a million utterances, spoken in En- glish, from over 7,000 speakers. Furthermore, the two datasets have a complicated and well-designed data collection pipeline, thus can represent ’real world’ situations. Using VoxCeleb2 as training data, the author …

Content truncated to 500 words