All projects

Research · 2022 · Research at KJ Somaiya College of Engineering · IEEE ICAST 2022

Dysarthria Detection from Speech

Can a network hear a motor disorder in 25 milliseconds of speech?

A hybrid CNN-LSTM that classifies dysarthric speech from 128 MFCC features, reaching 99.59% accuracy on the TORGO recordings of speakers with cerebral palsy or ALS. Published at IEEE ICAST 2022.

IEEE Xplore
speech · 25 ms frames128 MFCC features4 conv blocksLSTM · 256 unitsDysarthriaNon-dysarthria99.59%TORGO · 9,184 recordings · 5-fold 99.90%

Problem

Dysarthria, a motor speech disorder, is assessed by a speech-language pathologist one patient at a time. A screening model could take the first pass.

Approach

  • TORGO: 9,184 recordings from speakers with and without dysarthria (cerebral palsy or ALS), augmented by speed and loudness to 18,366 clips.
  • 128 MFCC features per clip: 25 ms frames, pre-emphasis, Hamming window, FFT, Mel log spectrogram, DCT.
  • A 25-layer hybrid: four convolutional blocks as the feature extractor, an LSTM with 256 units as the classifier.
  • Compared against CNN-SVM, random forest, decision tree and XGBoost, with five-fold cross-validation.

Stack

  • Python
  • TensorFlow
  • Keras
  • Librosa

Outcome

  • 99.59% test accuracy; 99.90% mean over five folds
  • Ahead of the published transfer-learning result on the same data (97.73%)

Figures

framing · 25 mspre-emphasisHamming windowFFTMel log spectrogramDCT128 MFCC featuresEach clip becomes a 128-value fingerprint before the network sees it.
From waveform to 128 MFCC features.
predictednon-dysarthriadysarthriaactual non-dysarthriaactual dysarthria226846161344trained on the original recordings, tested on 3,674 augmented clips62 errors in 3,674
Confusion matrix on the 3,674-clip augmented test set: 62 errors.