An Attention-Enhanced CNN-BiLSTM Model for Speech Assessment of Hard-of-Hearing Speakers
IEEE Access · 2026 · 14 · 46916-46931
DOI: 10.1109/access.2026.3673713 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Sandeep Chandran und 4 weiteren Autor:innen. Die Arbeit nennt 52 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Centre of Excellence of Disability Studies (CEDS), Kerala.
Abstract (Originalsprache)
Speech recognition systems work well for people who can hear normally; however, recognizing speech from hard-of-hearing speakers remains challenging due to limited datasets and distinct articulation characteristics. The proposed model uses a hybrid deep learning model using a convolutional neural network, a bi-directional long short–term memory network, and an attention mechanism for HoH speech recognition. Log–Mel spectrograms, along with their first-order (delta) and second–order (delta–delta) derivatives, are employed for the input features to represent spectral dynamics and the changes over time. To test the proposed method, a HoH speech dataset was developed with seventeen words taken from the standard word list developed by the All India Institute of Speech and Hearing, Mysore. Ten volunteers from the National Institute of Speech and Hearing, Kerala, participated in dataset creation. The model was further validated on the public Google Speech Commands dataset due to the absence of a standard HoH benchmark. The result of the experiment shows that the proposed model has an average accuracy of 95.20% and an F1-score of 0.95 for recognizing HoH speech effectively. The results show that the inclusion of temporal derivative features into an attention-based CNN-BiLSTM model results in a robust solution for assistive speech recognition systems.
Förderung
- Centre of Excellence of Disability Studies (CEDS), Kerala
Zitieren
APA
Chandran, S., Mohan, J., Sreejith, S., Raji, N. R., & Anjana, A. V. (2026). An Attention-Enhanced CNN-BiLSTM Model for Speech Assessment of Hard-of-Hearing Speakers. IEEE Access, 14, 46916-46931. https://doi.org/10.1109/access.2026.3673713
Vancouver
Chandran S, Mohan J, Sreejith S, Raji NR, Anjana AV. An Attention-Enhanced CNN-BiLSTM Model for Speech Assessment of Hard-of-Hearing Speakers. IEEE Access. 2026;14:46916-46931. doi:10.1109/access.2026.3673713
BibTeX
@article{Chandran2026,
author = {Chandran, Sandeep and Mohan, Jesna and Sreejith, S. and Raji, N. R. and Anjana, A. V.},
title = {An Attention-Enhanced CNN-BiLSTM Model for Speech Assessment of Hard-of-Hearing Speakers},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {46916--46931},
doi = {10.1109/access.2026.3673713}
}
EUVIMED