Robust Dysarthric Speech Transcription via Transformer-Based Whisper ASR: Spectral-Temporal Modeling for Impaired Articulation

Automatic transcription of dysarthric speech remains a significant challenge due to slurred articulation, phonetic distortions, and variability in speech clarity caused by neuromuscular impairments. In this study, we leverage OpenAI’s Whisper, an encoder–decoder ASR model, to transcribe dysarthric speech from the TORGO dataset, using a carefully selected subset of 100 audio files (50 dysarthric and 50 normal speech recordings), forming 49-word pairs for evaluation. Audio recordings were preprocessed to standardize sampling rate and format, and speech representations were extracted using log-Mel spectrograms, enabling robust representation of spectral and temporal patterns despite impaired articulation. The proposed Whisper model achieved an average Word Error Rate (WER) of 1.30 errors per word, with substitution errors dominating, followed by deletion and insertion errors. Variability analyses (box plots and WER histograms) demonstrate consistent transcription performance across different dysarthric speech samples. Words with clearer articulation or prolonged phonation were transcribed more accurately, while severely distorted words contributed to higher error rates. These results provide strong quantitative evidence of Whisper’s robustness, demonstrating its capability to handle a wide range of dysarthric speech patterns and establishing its effectiveness as a reliable tool for dysarthric speech recognition in real-world ASR applications.