Improving dysarthric speech recognition using empirical mode decomposition and convolutional neural network
Problem
Automatic speech recognizers perform poorly on dysarthric speech, making voice technology unreliable for affected speakers.
Why it matters
Accurate recognition of impaired speech is essential for accessible voice technology for people with dysarthria.
What the AI did
EMD with Hurst-based mode selection enhances dysarthric speech as preprocessing; MFCC features then feed a CNN-based recognizer, improving accuracy over baseline systems.
Results
Overall accuracy increased by 20.72% over HMM-GMM baselines and 9.95% over a CNN without enhancement, using k-fold cross-validation on the Nemours corpus.
Impact
Demonstrated accuracy gains on the Nemours corpus; real-world deployment not stated.
Record
| Venue | EURASIP Journal on Audio Speech and Music Processing |
|---|---|
| Published | 2020 |
| Work type | article |
| Authors | Mohammed Sidi Yakoub, Sid‐Ahmed Selouani, Brahim-Fares Zaidi, Asma Bouchair |
| Citations | 65 (OpenAlex, at import time) |
| References | 17 |
| Model or method | Convolutional neural network (CNN) |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1186/s13636-019-0169-5 |
| Open access | gold |
| Keywords | dysarthric speech, speech recognition, CNN, empirical mode decomposition, Nemours corpus |
Importance rating
real but narrower advance
Dysarthric speech recognition enables accessible voice technology for affected speakers; gains shown only on Nemours corpus, no population or deployment scale stated.