Research on Robustness of Voiceprint Recognition Technology

Shen, H.; Wang, Bo; Wang, Junsheng

doi:10.1145/3302425.3302467

Cited by 4 publications

(2 citation statements)

References 4 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Snyder D et al put forward a new acoustic feature Xvector [19], which maps the variable-length speech signal into a fixed-dimensional space through DNN, and this method can make full use of the training data compared with i-vector technology. Chung J et al collected the largest voiceprint recognition data set from opensource media at present, and can identify the identity in speech efficiently under various conditions through Convolutional Neural Network (CNN), especially under noisy and unrestricted conditions [20]; In the research of voiceprint feature denoising method, Shen H et al adopted the endpoint detection algorithm based on spectrum and used Empirical Mode Decomposition (EMD) algorithm to reconstruct the spectrum, which can effectively reduce voiceprint noise [21]. In addition, Wan L et al proposed a new loss function based on the deep neural network [22], called Generalized End-To-End (GE2E) loss, which makes the training of the speaker verification model more effective than the previous Tuple-Based End-To-End (TE2E) loss function.…”

Section: Research Status Of Voiceprint Recognitionmentioning

confidence: 99%

Deep Learning-based Analysis of Voiceprint Data Mining

Tang¹

2022

JSER

View full text Add to dashboard Cite

In the information age, the intelligent data mining method represented by deep learning is playing an important role in various fields at present. It is necessary to study how to efficiently use the intelligent data mining method to obtain valuable information from massive information. Open-set voiceprint recognition is realized by intelligent data mining technology. Therefore, it is of great practical significance to achieve rapid and accurate identification of the speaker's identity. Because the traditional voiceprint recognition method has insufficient ability to distinguish the speakers inside and outside the set, it often leads to a high false recognition rate. Mining parameters containing more speakers' personality characteristics and how to calculate the threshold become the bottleneck problems of open set voiceprint recognition. Therefore, this paper adopts the deep confidence network stacked by three layers of restricted Boltzmann machines as the deep acoustic feature extractor. The mel-frequency cepstral coefficients of 24-dimensional basic acoustic features are mapped to 256dimensional feature space, and the parameters of deep acoustic features containing more speaker's personality characteristics are obtained. Then, an open-set adaptive threshold calculation algorithm is obtained. In this paper, the similarity value of deep acoustic features is calculated by the Gaussian mixture model, and the maximum inter-class variance of the similarity value is calculated by the OTSU algorithm. When the inter-class variance is the maximum, the similarity value is the best threshold. The experimental test shows that the algorithm for calculating threshold based on deep learning proposed in this paper has a lower false rejection rate and lower false rejection rate.

show abstract

Section: Research Status Of Voiceprint Recognitionmentioning

confidence: 99%

Deep Learning-based Analysis of Voiceprint Data Mining

Tang¹

2022

JSER

View full text Add to dashboard Cite

show abstract

“…Robustness is becoming crucial to neural networks with the successful application of deep learning to safety-critical domains such as self-driving (Gopinath et al 2018) and access control by face and voiceprint recognition (Goswami et al 2018;Shen et al 2018). AI systems powered by non-robust neural networks are vulnerable and fragile to the perturbation from the environment and adversarial attack (Moosavi-Dezfooli, Fawzi, and Frossard 2016;Fawzi et al 2017).…”

Section: Introductionmentioning

confidence: 99%

Tightening Robustness Verification of Convolutional Neural Networks with Fine-Grained Linear Approximation

Zhang

2021

AAAI

View full text Add to dashboard Cite

The robustness of neural networks can be quantitatively indicated by a lower bound within which any perturbation does not alter the original input’s classification result. A certified lower bound is also a criterion to evaluate the performance of robustness verification approaches. In this paper, we present a tighter linear approximation approach for the robustness verification of Convolutional Neural Networks (CNNs). By the tighter approximation, we can tighten the robustness verification of CNNs, i.e., proving they are robust within a larger 10 perturbation distance. Furthermore, our approach is applicable to general sigmoid-like activation functions. We implement DeepCert, the resulting verification toolkit. We evaluate it with open-source benchmarks, including LeNet and the models trained on MNIST and CIFAR. Experimental results show that DeepCert outperforms other state-of-the-art robustness verification tools with at most 286.28% improvement to the certified lower bound and 1566.76 times speedup for the same neural networks.

show abstract