Real Additive Margin Softmax for Speaker Verification

Li, Lantian; Nai, Ruiqian; Wang, Dong

doi:10.1109/icassp43922.2022.9747166

Cited by 14 publications

(5 citation statements)

References 19 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The second is to add an additional trainable neural network module, e.g., decision residual networks (Dr-vectors) [21], deep learning backend (DLB) [22] and tied variational autoencoder (TVAE) [23]. And the last is to develop a robust backend against domain mismatch, such as Coral++ [24], domain-aware batch normalization (DABN) and domain-agnosticinstance normalization (DAIN) [25], information-maximized variational domain adversarial neural network (InfoVDANN) [26], and etc. However, these algorithms rarely use spatial or graph information among the extracted embeddings, which may significantly boost the performance.…”

Section: Introductionmentioning

confidence: 99%

Crossed-Time Delay Neural Network for Speaker Recognition

Chen

Liang

Shi

et al. 2021

MultiMedia Modeling

View full text Add to dashboard Cite

Time Delay Neural Network (TDNN) is a well-performing structure for DNN-based speaker recognition systems. In this paper we introduce a novel structure Crossed-Time Delay Neural Network (CTDNN) to enhance the performance of current TDNN. Inspired by the multi-filters setting of convolution layer from convolution neural network, we set multiple time delay units each with different context size at the bottom layer and construct a multilayer parallel network. The proposed CTDNN gives significant improvements over original TDNN on both speaker verification and identification tasks. It outperforms in VoxCeleb1 dataset in verification experiment with a 2.6% absolute Equal Error Rate improvement. In few shots condition CTDNN reaches 90.4% identification accuracy, which doubles the identification accuracy of original TDNN. We also compare the proposed CTDNN with another new variant of TDNN, FTDNN, which shows that our model has a 36% absolute identification accuracy improvement under few shots condition and can better handle training of a larger batch in a shorter training time, which better utilize the calculation resources.

show abstract

Section: Introductionmentioning

confidence: 99%

Crossed-Time Delay Neural Network for Speaker Recognition

Chen

Liang

Shi

et al. 2021

MultiMedia Modeling

View full text Add to dashboard Cite

show abstract

“…AAMSoftmax [16] is most commonly used as a loss function in speaker verification tasks, and it can be expressed as (1):…”

Section: B Adaptive Loss Functions In Speaker Verification Tasksmentioning

confidence: 99%

AWLloss: Speaker Verification Based on the Quality and Difficulty of Speech

Liu¹

2023

Preprint

View full text Add to dashboard Cite

<p>Speaker verification is a natural and effective biometric authentication method. This method may be more practical when trained using data from real, unconstrained scenarios. However, although the speech quality from real scenarios varies because of various adverse factors, including extreme noise and the lack of identity information, most models treat all speech samples equally. To tackle this deficiency, we propose adaptive weight loss (AWL), a function that assigns different weights to samples in accordance with their quality and difficulty based on our finding that speech quality is positively correlated with the L2 norm of speaker embedding. AWL directs the model's attention to high-quality difficult speech samples and ignores low-quality difficult samples as much as possible so that the model can learn well from speech data containing a lot of noise. Comprehensive experiments reveal that AWL significantly outperforms existing loss functions on both utilized public and noisy datasets.</p>

show abstract

“…Then, 80-dimensional filter bank is extracted as acoustic features using 25ms window length and 10ms hop size. Plus, loss function is AAM-softmax [24] with a 0.2 margin and a scale of 32. ResNets adopt SGD optimizer with a momentum of 0.9 and a weight decay of 1e-4, while DF-ResNets are optimized using AdamW [68] with a weight decay of 0.05.…”

Section: Training Strategiesmentioning

confidence: 99%

“…Lastly, the whole system is trained using a multi-class speaker classifier. To enhance systems' performance, researchers have exerted numerous endeavors in multiple aspects, such as network backbones [4]- [15], pooling strategies [16]- [20], and training criteria [21]- [24].…”

Section: Introductionmentioning

confidence: 99%