Scholarlink

<p>In recent years, the Transformer model has made significant progress in the lightweight direction of speech recognition. However, the process of lightweight is often accompanied by a loss of accuracy. In this paper, Low-Rank Adaptive Transformer (LRAT) is proposed. Transformer is used as the backbone model to improve the ability of context information extraction by introducing Attention-Enhanced VGG (AE-VGG) in the feature extraction stage. Low-Rank Adaptive Multi-Head Attention is introduced into encoders and decoders to reduce computational complexity. Experiments were conducted using the AiShell-1, THCHS-30, and LibriSpeech datasets. On the AiShell-1 dataset, our model reduced the character error rate (CER) by 1.70% and 1.30% compared to the Transformer and Low-Rank Transformer, respectively, and reduced the parameters by 13.29M and 0.89M, respectively. On the THCHS-30 dataset, the CER was reduced by 0.64% and 1.69%, respectively, and the parameters were reduced by 10.8M and -1.1M, respectively. On the LibriSpeech dataset, the WER was reduced by 0.08% and 0.14%, respectively, and the parameters were reduced by 16.44M and 0.79M, respectively. The experimental results show that LRAT achieves a good balance between lightweight and accuracy.</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

LRAT: Low-Rank Adaptive Transformer for end-to-end speech recognition

  • 期刊论文

  • 发表日期：

    2025年7月5日

查看全文
  • Xinmin Cheng,
  • Yuyang Wu,
  • Ruiqin Wang

摘要

In recent years, the Transformer model has made significant progress in the lightweight direction of speech recognition. However, the process of lightweight is often accompanied by a loss of accuracy. In this paper, Low-Rank Adaptive Transformer (LRAT) is proposed. Transformer is used as the backbone model to improve the ability of context information extraction by introducing Attention-Enhanced VGG (AE-VGG) in the feature extraction stage. Low-Rank Adaptive Multi-Head Attention is introduced into encoders and decoders to reduce computational complexity. Experiments were conducted using the AiShell-1, THCHS-30, and LibriSpeech datasets. On the AiShell-1 dataset, our model reduced the character error rate (CER) by 1.70% and 1.30% compared to the Transformer and Low-Rank Transformer, respectively, and reduced the parameters by 13.29M and 0.89M, respectively. On the THCHS-30 dataset, the CER was reduced by 0.64% and 1.69%, respectively, and the parameters were reduced by 10.8M and -1.1M, respectively. On the LibriSpeech dataset, the WER was reduced by 0.08% and 0.14%, respectively, and the parameters were reduced by 16.44M and 0.79M, respectively. The experimental results show that LRAT achieves a good balance between lightweight and accuracy.