Lihong Deng

Lihong Deng

Auto-generated profile, based on publicly available information

Lihong Deng

School of Computer and Artificial Intelligence, Southwest Jiaotong University, Chengdu, China

Research Focus & Metrics

Academic Career Length

4Years (2023-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
20 January 2023 in Sensors
Verified
Journal Article
Open access
PDF file
FDFei Deng
Fei Deng
LDLihong Deng
Lihong Deng
PJPeifan Jiang
Peifan Jiang
Gexiang ZhangGexiang Zhang
Gexiang Zhang
QYQiang Yang
Qiang Yang
In speaker recognition tasks, convolutional neural network (CNN)-based approaches have shown significant success. Modeling the long-term contexts and efficiently aggregating the information are two challenges in speaker recognition, and they have a critical impact on system performance. Previous research has addressed these issues by introducing deeper, wider, and more complex network architectures and aggregation methods. However, it is difficult to significantly improve the performance with these approaches because they also have trouble fully utilizing global information, channel information, and time-frequency information. To address the above issues, we propose a lighter and more efficient CNN-based end-to-end speaker recognition architecture, ResSKNet-SSDP. ResSKNet-SSDP consists of a residual selective kernel network (ResSKNet) and self-attentive standard deviation pooling (SSDP). ResSKNet can capture long-term contexts, neighboring information, and global information, thus extracting a more informative frame-level. SSDP can capture short- and long-term changes in frame-level features, aggregating the variable-length frame-level features into fixed-length, more distinctive utterance-level features. Extensive comparison experiments were performed on two popular public speaker recognition datasets, Voxceleb and CN-Celeb, with current state-of-the-art speaker recognition systems and achieved the lowest EER/DCF of 2.33%/0.2298, 2.44%/0.2559, 4.10%/0.3502, and 12.28%/0.5051. Compared with the lightest x-vector, our designed ResSKNet-SSDP has 3.1 M fewer parameters and 31.6 ms less inference time, but 35.1% better performance. The results show that ResSKNet-SSDP significantly outperforms the current state-of-the-art speaker recognition architectures on all test sets and is an end-to-end architecture with fewer parameters and higher efficiency for applications in realistic situations. The ablation experiments further show that our proposed approaches also provide significant improvements over previous methods.

15 November 2023 in Engineering Applications of Artificial Intelligence
Verified
Journal Article
Closed access
LDLihong Deng
Lihong Deng
FDFei Deng
Fei Deng
KZKepeng Zhou
Kepeng Zhou
PJPeifan Jiang
Peifan Jiang
GZGexiang Zhang
Gexiang Zhang
QYQiang Yang
Qiang Yang

17 August 2023 in IEEE Access
Verified
Journal Article
Open access
JWJiji Wang
Jiji Wang
FDFei Deng
Fei Deng
LDLihong Deng
Lihong Deng
PGPing Gao
Ping Gao
YHYuanxiang Huang
Yuanxiang Huang
Current speaker recognition systems are mainly for the combined application of network architectures and attention mechanisms, however, lightweight networks are not able to extract frame-level features of speaker speech well, and deeper and wider networks also face the problems of slower inference and excessive number of parameters. To this end, we proposes Split-ResNet, a network structure for split residuals, which can obtain a combination of multiple receptive field at a finer-grained level, thus obtaining a variety of feature representations with different scale combinations and producing more informative and comprehensive multi-scale features. In addition we propose a dual time-frequency attention (DTFA) that enhances key features and suppresses unimportant features by focusing on features in the time and frequency domains and learning weights from the time and frequency channels, respectively. We finally tested the speaker recognition system using a combination of Split-ResNet and DTFA against other speaker recognition systems on the Voxceleb1-O test set. The test results show that the speaker recognition system proposed in this paper is 0.98%, 0.39%, 0.69% and 0.47% lower in EER compared with SpeechNAS, RawNet2, Y-vector and CNN+Transformer, respectively, proving that DTFA+Split-ResNet is a speaker recognition system with good speaker audio feature extraction capability and discriminative capability.

27 April 2026 in IEEE Signal Processing Letters
Verified
Journal Article
Closed access
FDFei Deng
Fei Deng
LDLihong Deng
Lihong Deng
PJPeifan Jiang
Peifan Jiang
Systems based on Time Delay Neural Networks (TDNNs) have achieved remarkable success in speaker verification due to their favorable balance between accuracy and computational cost. However, under short-utterance conditions, two limitations arise: limited information hinders the extraction of discriminative speaker features, and the inherent one-dimensional (1D) convolutions lack translation equivariance to frequency shifts, restricting robustness and feature utilization. To address these issues, we propose SRC-TDNN (Split Residual Collaborative TDNN), a lightweight architecture that integrates a two-dimensional (2D) convolutional component with a TDNN component. The 2D convolutional component employs parallel split residual convolutional (SplitResConv) blocks to capture multi-scale time–frequency features, followed by a Fourier Fusion Module (FFM) to enhance feature discriminability. By sharing weights along the frequency axis, this component achieves translation equivariance, alleviating the impact of frequency shifts in short utterances. The TDNN component further aggregates the fused features along the temporal axis while inheriting the translation equivariance from the 2D component. It also adopts parallel 1D SplitResConv blocks to capture both short-term phonetic details and long-term speaker characteristics across multiple temporal scales. A multi-branch architecture is employed throughout, enabling both network-level ( macro) and block-level (micro) multi-scale modeling to better exploit limited information. Evaluations on four test sets with varying utterance lengths demonstrate that the proposed method achieves the best overall performance, particularly under short-utterance conditions. Compared with ECAPA-TDNN, SRC-TDNN reduces the number of parameters by only 0.4 M while achieving at least a 30% relative improvement.