Scholarlink

<p>This study introduces an end-to-end fusion network, SE-ResNeXt, incorporating a channel attention mechanism residual network and a Long Short-Term Memory network (LSTM) for enhancing speaker recognition in noisy environments. Spectrogram decomposition through non‑negative matrix factorization (NMF) yields two matrices: the basis matrix W and the coefficient matrix H. These are used as inputs to the fusion network. These matrices are processed in parallel by the SE-ResNeXt network and LSTM network. The training and testing utilized speech data from 200 speakers in the AISHELL-ASR0009-OS1 dataset. The experiment assessed the network's robustness by introducing three common background noises. Results indicate that the proposed network achieved an average recognition accuracy of 95.6% in the absence of noise. Under various SNR conditions, the model demonstrates superior noise robustness and generalization compared to baseline models. Notably, when subjected to “Leopard” noise at an SNR of 5, the model maintained a test accuracy of 90.7%. In comparison with CNN, CNN + GRU, and CNN + Residual models, the proposed model exhibited performance improvements of 34.9%, 65.7%, and 52.3%, respectively.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A study on speaker recognition based on improved SE-ResNeXt network

  • 期刊论文

  • 发表日期：

    2026年3月14日

查看全文
  • Dongbo Liu,
  • Sitian Wang,
  • Liming Huang,
  • Yu Fang

摘要

This study introduces an end-to-end fusion network, SE-ResNeXt, incorporating a channel attention mechanism residual network and a Long Short-Term Memory network (LSTM) for enhancing speaker recognition in noisy environments. Spectrogram decomposition through non‑negative matrix factorization (NMF) yields two matrices: the basis matrix W and the coefficient matrix H. These are used as inputs to the fusion network. These matrices are processed in parallel by the SE-ResNeXt network and LSTM network. The training and testing utilized speech data from 200 speakers in the AISHELL-ASR0009-OS1 dataset. The experiment assessed the network's robustness by introducing three common background noises. Results indicate that the proposed network achieved an average recognition accuracy of 95.6% in the absence of noise. Under various SNR conditions, the model demonstrates superior noise robustness and generalization compared to baseline models. Notably, when subjected to “Leopard” noise at an SNR of 5, the model maintained a test accuracy of 90.7%. In comparison with CNN, CNN + GRU, and CNN + Residual models, the proposed model exhibited performance improvements of 34.9%, 65.7%, and 52.3%, respectively.