Scholarlink

<p>Dysarthria is a motor speech disorder that impairs articulation and intelligibility, making communication challenging. Keyword spotting (KWS) offers valuable support for dysarthric speakers by enabling voice-activated control where continuous ASR systems often fail. However, KWS for dysarthric speech remains underexplored. This study introduces a new framework that utilizes self-supervised learning (SSL) features to enhance KWS performance in dysarthric speech. We extract feature embeddings from the final encoder layers of three state of art SSL models Wav2Vec2, HuBERT, and Data2Vec and integrate these into a DNN-based KWS system. We use TORGO database and choose leave-one-speaker-out (LOSO) evaluation setup and compare the performance of SSL-based features against traditional Kaldi-derived features: MFCC, PLP and filter-bank (F-Bank). SSL-based features outperform traditional ones, with HuBERT achieving the best performance with an average ATWV score of 0.3295 for 10 keywords and 0.3573 for 20 keywords, improving over PLP based features which performed the best among the traditional Kaldi features. To enhance the generalization of the framework we conducted all the experiments with additional dataset, UASpeech and get similar trends. HuBERT outperformed the traditional features with an ATWV of 0.3732 and 0.3995 for 10 and 20 keywords. Statistical analysis confirms these gains are robust: paired t-tests show highly significant improvements (p &lt; 0.001) for both low and high severity speakers, and Wilcoxon signed-rank tests (p = 0.0156) further support these results. Moreover, 95% confidence intervals demonstrate consistent performance boosts, highlighting the reliability of SSL embeddings over traditional Kaldi features. We also performed pearson correlation to assess the consistency of the features. These findings highlight the potential of SSL-based features to advance dysarthric keyword spotting system.</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

Empowering dysarthric communication: a self-supervised approach to keyword spotting

  • 期刊论文

  • 发表日期：

    2025年12月9日

查看全文
  • Subham Kutum,
  • Paban Sapkota,
  • Abhijit Sinha,
  • Hemant Kumar Kathania,
  • Mahesh Chandra Govil

摘要

Dysarthria is a motor speech disorder that impairs articulation and intelligibility, making communication challenging. Keyword spotting (KWS) offers valuable support for dysarthric speakers by enabling voice-activated control where continuous ASR systems often fail. However, KWS for dysarthric speech remains underexplored. This study introduces a new framework that utilizes self-supervised learning (SSL) features to enhance KWS performance in dysarthric speech. We extract feature embeddings from the final encoder layers of three state of art SSL models Wav2Vec2, HuBERT, and Data2Vec and integrate these into a DNN-based KWS system. We use TORGO database and choose leave-one-speaker-out (LOSO) evaluation setup and compare the performance of SSL-based features against traditional Kaldi-derived features: MFCC, PLP and filter-bank (F-Bank). SSL-based features outperform traditional ones, with HuBERT achieving the best performance with an average ATWV score of 0.3295 for 10 keywords and 0.3573 for 20 keywords, improving over PLP based features which performed the best among the traditional Kaldi features. To enhance the generalization of the framework we conducted all the experiments with additional dataset, UASpeech and get similar trends. HuBERT outperformed the traditional features with an ATWV of 0.3732 and 0.3995 for 10 and 20 keywords. Statistical analysis confirms these gains are robust: paired t-tests show highly significant improvements (p < 0.001) for both low and high severity speakers, and Wilcoxon signed-rank tests (p = 0.0156) further support these results. Moreover, 95% confidence intervals demonstrate consistent performance boosts, highlighting the reliability of SSL embeddings over traditional Kaldi features. We also performed pearson correlation to assess the consistency of the features. These findings highlight the potential of SSL-based features to advance dysarthric keyword spotting system.