Automatic Classification of Speech Dysarthric Intelligibility Levels Using Textual Feature

The comprehension of human language is fundamentally important in modern intelligent systems. Automatic Speech Intelligibility assessment involves determining the efficiency with which another speaker comprehends the same speech. Changes in the acoustic perception of speech may be due to neurological disorders or accidental ingestion of toxic substances. Dysarthria is one such disorder that affects the pronunciation of speech due to deterioration in the neuromuscular control of speech organs. Dysarthric speech is characterized by impairments that hinder clear pronunciation and understanding, presenting significant challenges for accurate transcription and classification. To the best of our knowledge, no previous studies have focused exclusively on using textual features for this purpose, with most existing research relying primarily on speech data. This study introduces a novel Hybrid CNN-BLSTM model that utilizes textual features extracted through Automatic Speech Recognition (ASR) to classify intelligibility levels in dysarthric speech using the TORGO dataset. Our approach addresses transcription inconsistencies by employing ASR to convert speech to text. We compare the proposed Hybrid CNN-BLSTM model with state-of-the-art deep learning models, including Long Short-Term Memory (LSTM), Bidirectional LSTM (BLSTM), Gated Recurrent Unit (GRU), and Convolutional Neural Network (CNN). Experimentally, the Hybrid CNN-BLSTM model, optimized with the Stochastic Gradient Descent (SGD), achieved the highest accuracy, reaching 90 % in classifying intelligibility levels.