An Effective Deep Embedding Learning Method Based on Dense-Residual Networks for Speaker Verification

Abstract

In this paper, we present an effective end-to-end deep embedding learning method based on Dense-Residual networks, which combine the advantages of a densely connected convolutional network (DenseNet) and a residual network (ResNet), for speaker verification (SV). Unlike a model ensemble strategy which merges the results of multiple systems, the proposed Dense-Residual networks perform feature fusion on every basic DenseR building block. Specifically, two types of DenseR blocks are designed. A sequential-DenseR block is constructed by densely connecting stacked basic units in a residual block of ResNet. A parallel-DenseR comprises split and concatenation operations on residual and dense components via corresponding skip connections. These building blocks are stacked into deep networks to exploit the complementary information with different receptive field sizes and growth rates. Extensive experiments have been conducted on the VoxCeleb1 dataset to evaluate the proposed methods. The SV performance achieved by the proposed Dense-Residual networks is shown to outperform corresponding ResNet, DenseNet or fusions of them, with similar model complexity, by a significant margin.

Few-Shot Speaker Identification Using Lightweight Prototypical Network With Feature Grouping and Interaction

“…These features mainly included the X-vector learned by a Time-Delay Neural Network (TDNN) [19]- [23] or an Emphasized Channel Attention, Propagation and Aggregation in TDNN (ECAPA-TDNN) [24]; the R-vector learned by a Residual Network with 34 layers (ResNet34) [25]; the S-vector learned by a Transformer [26]. In addition, other kinds of neural networks were adopted to learn deep embeddings [27]- [35], such as temporal dynamic convolutional neural network [31], Attentive Multi-scale Convolutional Recurrent Network (AMCRN) [33], Siamese neural network [34], and long short-term memory network [35].…”

MFFA: Music Feature Fusion Attention Model for Music Emotion Classification

“…For example, Gupta et al [14] proposed a residual connection between input and the output of the inception block to have a continuity between the previously observed pose and the next predicted pose. Combined the advantages of DenseNet and ResNet with dense connection, Liu et al [15] designed sequential Dense-Residual blocks and parallel Dense-Residual blocks for feature fusion. A multimodal emotion recognition model based on feature fusion and residual connection was proposed by Du et al [16], which uses Bi-LSTM and Muti-head attention to extract key features in voice, text, and video, and combined residual connection mechanism to prevent gradient from disappearing.…”

“…The main idea of attention mechanism is to introduce a dynamic weighting mechanism in the network that allows the model to perform different weighted calculations based on different parts of the input, thus enabling the network to focus more on important information and ignore irrelevant information [2,10,11]. On the other side, in a traditional deep neural network, it may lead to loss of information and degradation of the model as the depth of the network increases, so residual learning is introduced where multi-scale and multi-level information are integrated through residual connections [9,[14][15][16].…”