2026-03-16T11:34:25Z https://ipsj.ixsq.nii.ac.jp/oai

oai:ipsj.ixsq.nii.ac.jp:00231272 2025-01-19T10:49:43Z 1164:4179:11237:11430

Enhancing Dysarthric Speech Recognition with Auxiliary Feature Fusion Module: Exploring Articulatory-related Features from Foundation Models Enhancing Dysarthric Speech Recognition with Auxiliary Feature Fusion Module: Exploring Articulatory-related Features from Foundation Models Yuqin, Lin Longbiao, Wang Jianwu, Dang Nobuaki, Minematsu Yuqin, Lin Longbiao, Wang Jianwu, Dang Nobuaki, Minematsu ポスター Addressing dysarthric speech variability in Automatic Speech Recognition (ASR) is crucial for improving human-computer interactions for everyone. This paper proposes the Auxiliary Features Fusion (AFFusion) module, which leverages phonetic and articulatory-related features from models like wav2vec to compensate for distorted acoustics in dysarthric ASR. Experimental results using AFFusion with various feature models demonstrate its effectiveness on dysarthric databases. Interestingly, the analysis suggests that AFFusion shares similarities with human speech perception processes, offering potential insights into addressing fuzzy recognition in dysarthric ASR based on the motor theory of speech perception. Addressing dysarthric speech variability in Automatic Speech Recognition (ASR) is crucial for improving human-computer interactions for everyone. This paper proposes the Auxiliary Features Fusion (AFFusion) module, which leverages phonetic and articulatory-related features from models like wav2vec to compensate for distorted acoustics in dysarthric ASR. Experimental results using AFFusion with various feature models demonstrate its effectiveness on dysarthric databases. Interestingly, the analysis suggests that AFFusion shares similarities with human speech perception processes, offering potential insights into addressing fuzzy recognition in dysarthric ASR based on the motor theory of speech perception. 情報処理学会 2023-11-25 eng technical report https://ipsj.ixsq.nii.ac.jp/records/231272 2188-8779 AN10115061 研究報告自然言語処理（NL） 2023-NL-258 14 1 6 https://ipsj.ixsq.nii.ac.jp/record/231272/files/IPSJ-NL23258014.pdf application/pdf 1.7 MB