Dysarthric Speech Recognition using Convolutional Recurrent Neural Networks
Problem
Automatic speech recognition performs poorly on dysarthric (impaired) speech, which is unintelligible, irregular, and data-scarce.
Why it matters
Impaired speakers cannot use mainstream voice interfaces, limiting accessibility for people with motor speech disorders.
What the AI did
A deep convolutional recurrent neural network (CRNN) was developed and trained on the Torgo dataset of impaired and unimpaired speech, then compared to a vanilla CNN.
Results
CRNN attains 40.6% versus 31.4% for vanilla CNN on the Torgo dataset; evaluation metric not stated in source.
Impact
CRNN improved over the CNN baseline, but the reported score indicates dysarthric speech recognition remains far from solved.
Record
| Venue | International journal of intelligent engineering and systems |
|---|---|
| Published | 2020 |
| Work type | article |
| Authors | Hussain Albaqshi, Alaa Sagheer |
| Citations | 24 (OpenAlex, at import time) |
| References | 42 |
| Model or method | convolutional recurrent neural network (CRNN) |
| Organization | not stated in source |
| Primary source | https://doi.org/10.22266/ijies2020.1231.34 |
| Open access | bronze |
| Keywords | dysarthric speech, speech recognition, CRNN, deep learning, Torgo dataset, accessibility |
Importance rating
real but narrower advance
Dysarthric speech recognition failure blocks accessibility for people with motor speech disorders; abstract gives no scale figures, and results (40.6% vs 31.4%) show the problem remains far from solved.