Attention-Based CNN-BiGRU for Speech Clarity Estimation in ALS
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
This project presents an end-to-end system for automatic speech clarity estimation in patients with Amyotrophic Lateral Sclerosis (ALS). The proposed approach models speech clarity as a continuous regression problem, enabling more realistic and clinically meaningful monitoring of speech degradation compared to traditional classification-based methods.The system utilizes a hybrid deep learning architecture combining Convolutional Neural Networks (CNN) for spectral feature extraction, Bidirectional Gated Recurrent Units (BiGRU) for temporal modeling, and a self-attention mechanism to focus on informative speech segments. Additionally, a soft labeling strategy based on acoustic features such as spectral properties, harmonic-to-noise ratio (HNR), and voiced ratio is employed to better capture the continuous nature of speech clarity.The model is trained and evaluated on the TORGO dysarthric speech dataset, which contains real-world variability in speakers and recording conditions. Experimental results demonstrate strong performance, achieving a Pearson correlation of 0.778, Mean Absolute Error (MAE) of 0.083, and an R² score of 0.567, highlighting the effectiveness of the proposed method.To enable real-world usability, the model is integrated into a full-stack web application (Speech Clarity Analyser) that allows users to record speech, obtain clarity scores, and monitor changes over time. The application is built using React (frontend), FastAPI (backend), and Supabase (database and authentication), ensuring secure data handling and scalable deployment.This work contributes to the development of non-invasive, automated tools for clinical monitoring of speech impairment, with potential applications in early diagnosis, remote patient monitoring, and assistive healthcare technologies.
Fields of Study & Subjects
Voice and Speech Disorders
Speech Recognition and Synthesis
Speech and Audio Processing
Publication details
- Journal / source
- Zenodo (CERN European Organization for Nuclear Research)
- Publication date
- 27 March 2026
- Language
- EN
- Document type
- Preprint