Optimizing Neural Network Embeddings Using a Pair-Wise Loss for Text-Independent Speaker Verification

Dhamyal, Hira; Zhou, Tianyan; Raj, Bhiksha; Singh, Rita

doi:10.1109/asru46091.2019.9003794

Cited by 4 publications

(4 citation statements)

References 20 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The recent deep learning based speaker verification approaches can be primarily categorized into two main aspects: advanced network structure constructions [1,2,3,4,18] and effective loss function designs [6,19,20,21].…”

Section: Related Workmentioning

confidence: 99%

“…Logistic affinity loss [19] instead optimizes an end-to-end speaker verification model by building a learnable decision boundary to distinguish the similar pairs and dissimilar pairs. The quartet loss [21] explicitly computes a pair-wise distance loss in the embedding space and increases the gap between the similarity score distributions between the same class pairs and different class pairs. Self-adaptive soft voice activity detection (VAD) [31] incorporates a deep neural network based VAD into a deep speaker embedding system to reduce the domain mismatch.…”

Section: Related Workmentioning

confidence: 99%

See 1 more Smart Citation

SpeechNAS: Towards Better Trade-off between Latency and Accuracy for Large-Scale Speaker Verification

Zhu¹,

Kong²,

Lu³

et al. 2021

Preprint

View full text Add to dashboard Cite

Recently, has been a successful and popular approach for speaker verification, which employs a time delay neural network (TDNN) and statistics pooling to extract speaker characterizing embedding from variable-length utterances. Improvement upon the x-vector has been an active research area, and enormous neural networks have been elaborately designed based on the x-vector, e.g., extended TDNN (E-TDNN) [2], factorized TDNN (F-TDNN) [3], and densely connected TDNN (D-TDNN) [4]. In this work, we try to identify the optimal architectures from a TDNN based search space employing neural architecture search (NAS), named SpeechNAS. Leveraging the recent advances in the speaker recognition, such as high-order statistics pooling, multibranch mechanism, D-TDNN and angular additive margin softmax (AAM) loss with a minimum hyper-spherical energy (MHE), SpeechNAS automatically discovers five network architectures, from SpeechNAS-1 to SpeechNAS-5, of various numbers of parameters and GFLOPs on the large-scale text-independent speaker recognition dataset VoxCeleb1. Our derived best neural network achieves an equal error rate (EER) of 1.02% on the standard test set of VoxCeleb1, which surpasses previous TDNN based state-of-the-art approaches by a large margin.

show abstract

Section: Related Workmentioning

confidence: 99%

Section: Related Workmentioning

confidence: 99%

SpeechNAS: Towards Better Trade-off between Latency and Accuracy for Large-Scale Speaker Verification

Zhu¹,

Kong²,

Lu³

et al. 2021

Preprint

View full text Add to dashboard Cite

show abstract

“…I-vectors are statistical low-dimensional representations over the distributions of spectral features, and are commonly used in state-of-the-art speaker recognition systems [31] and age estimation systems [32], [33]. Respectively, 400-dimensional and 600-dimensional i-vectors are extracted for Fisher and SRE datasets using the state-of-the-art speaker identification system [34].…”

Section: A Datamentioning

confidence: 99%

Neural Regression Trees

Memon

Zhao

Raj

et al. 2019

2019 International Joint Conference on Neural Networks (IJCNN)

Self Cite

View full text Add to dashboard Cite

Regression-via-Classification (RvC) is the process of converting a regression problem to a classification one. Current approaches for RvC use ad-hoc discretization strategies and are suboptimal. We propose a neural regression tree model for RvC. In this model, we employ a joint optimization framework where we learn optimal discretization thresholds while simultaneously optimizing the features for each node in the tree. We empirically show the validity of our model by testing it on two challenging regression tasks where we establish the state of the art.

show abstract

“…In the training process of CNN, depth metric learning such as pair-wise loss, 9,10 triplet loss, 11,12 and n-pair loss. 13 Above methods are especially applied to FGIR to extract more accurate features on the tiny differ-ences between objects.…”

Section: Introductionmentioning

confidence: 99%

Centralized contrastive loss with weakly supervised progressive feature extraction for fine‐grained common thorax disease retrieval in chest x‐ray

et al. 2023

View full text Add to dashboard Cite

Background Medical images have already become an essential tool for the diagnosis of many diseases. Thus a large number of medical images are being generated due to the daily routine inspection. An efficient image‐based disease retrieval system will not only make full use of existing data, but also help physicians to prognosis the diseases. Medical image retrieval is represented by the classification and localization of common thorax diseases in x‐ray images. Although extensive efforts have been put into this field, there are still many challenges. Purpose Most of the existing fine‐grained image research methods just apply existing deep learning frameworks in extracting the image features. However, these high‐level features mainly focus on the global representations of the object, rather than simultaneously considering the local ones. It requires fine‐grained details to classify the images with similar lesion areas. Thus, it is necessary to combine the global features and local ones to make the features more discriminative. On the other hand, training CNN models based on current existing strategies have a high time complexity, and is hard to get the discriminative features mentioned above. In addition, the visual retrieval method of fine‐grained medical images still has the problem of insufficient sample data with accurate annotation information. Methods To address above challenges, we introduced a novel fine‐grained medical images retrieval method. First, a centralized contrastive loss (CCLoss) is proposed as our metric learning loss function. Parameters are updated by using the center point, which not only improves the distinguishing performance of features, but also effectively reduces the time complexity of the algorithm. In addition, a weakly supervised progressive feature extraction method is proposed to gradually extract the combined features. And the attention mechanism module is applied to screen the target information after the initial positioning for fine refinement, so as to separate the features with a high degree of discrimination. The retrieval of 14 different chest diseases is evaluated on the chest x‐ray datasets. Results Compared with the existing research methods, the proposed method shows a better retrieval result for Recall@8 by 2.26%∼4.6%$\%{\sim }4.6\%$ and achieves a very efficient training speed which is 100 times faster than the pair‐wise loss‐based training strategy. We also assessed the effects of Recall@k (k = 2, 4, 6, 8) for progressive features extracted from different steps to obtain a model with the best retrieval performance. Conclusions The proposed model is capable of learning discriminative representations from chest x‐ray datasets, and it achieves better performance compared with other state‐of‐the‐art methods. Therefore, the developed model would be useful in the diagnosis of common thorax disease or unknown chest disease.

show abstract

Optimizing Neural Network Embeddings Using a Pair-Wise Loss for Text-Independent Speaker Verification

Cited by 4 publications

References 20 publications

SpeechNAS: Towards Better Trade-off between Latency and Accuracy for Large-Scale Speaker Verification

SpeechNAS: Towards Better Trade-off between Latency and Accuracy for Large-Scale Speaker Verification

Neural Regression Trees

Centralized contrastive loss with weakly supervised progressive feature extraction for fine‐grained common thorax disease retrieval in chest x‐ray

Contact Info

Product

Resources

About