Key points are not available for this paper at this time.
Automatic speech recognition (ASR) is a valued tool for individuals with dysarthria, a speech impairment characterized by various pathological traits that differ from healthy speech. However, recognizing dysarthric speech, which is spoken by individuals with speech impairments, poses unique challenges due to its diverse characteristics such as rugged pronunciation, loudness that varies at different intervals, speech that has lot of delays, pauses that are inpredictable, excessive nasal sounds, explosive pronunciation, and airflow noise. The survey reveals the various models for dysarthric speech recognition. Deep learning technologies, unfurls an improved ASR performance leaps and bounds breaking the fluency and pronunciation barriers. Various feature extractions and identification of different types of dysarthria, including spastic, mixed, ataxic, hypokinetic, and hyperkinetic are explored. The performance of contemporary deep learning approaches in dysarthric speaker recognition (DSR) is tested using various datasets to determine accuracy. In conclusion the most effective DSR strategies are identified and areas for future investigation is suggested. However, speaker-dependent difficulties restrict the generalizability of acoustic models, and a lack of speech data impedes training on large datasets. The study throws light on how the effectiveness of ASR for dysarthric speech can be improved and further areas of research in the area are highlighted.
Antony et al. (Fri,) studied this question.