TL;DR
This paper explores how acoustic-phonetic features can be used to automatically assess dysarthric speech, aiming to predict intelligibility ratings and classify speakers as healthy or dysarthric. The study demonstrates that a carefully selected subset of these features can achieve high accuracy in both tasks, offering a promising avenue for more accessible and objective clinical assessments.
What Problem Does This Paper Address?
Dysarthria, a motor speech disorder resulting from neurological injury, significantly impacts speech production and intelligibility. Reduced intelligibility can lead to social isolation and a diminished quality of life. While speech therapy can improve intelligibility, its effectiveness needs to be assessed. Traditional methods rely on human listeners for transcription or rating, which are time-consuming and labor-intensive. While machine learning approaches like ASR and deep learning embeddings show promise, their direct link to specific speech deviations relevant for therapy can be unclear. This paper addresses the need for objective, interpretable, and automated methods to assess dysarthric speech by investigating the utility of acoustic-phonetic features for predicting intelligibility and classifying speech types.
Methods
The study employed a two-stage feature selection process followed by regression and classification analyses.
- Feature Selection:
- LASSO (Least Absolute Shrinkage and Selection Operator): Initially, LASSO was used to select a subset of relevant acoustic-phonetic features from a larger pool (103 features). This method helps prevent overfitting by penalizing less important features.
- Stepwise Linear Regression (GLM): Following LASSO, a stepwise linear regression approach was used to further refine the feature set for predicting intelligibility ratings. This method iteratively adds or removes features based on their statistical significance (using AIC criterion).
- Intelligibility Prediction:
- Stepwise Linear Regression: The features selected in the previous steps were used in a stepwise linear regression model to predict intelligibility scores (measured by Visual Analogue Scale - VAS).
- Speech Classification:
- Linear Discriminant Analysis (LDA): Features identified as significant in the regression model were then used for classification. LDA was applied in two ways:
- Subset-based LDA: LDA was applied to progressively larger subsets of features, ordered by their significance from the regression model, to observe how accuracy changes with feature inclusion.
- Pairwise LDA: All possible pairs of the selected features were exhaustively tested with LDA to identify the most effective combinations for classification.
- Leave-One-Speaker-Out (LOSO) Cross-validation: This rigorous cross-validation scheme was used for the LDA classification to ensure the model generalizes well to unseen speakers.
Datasets
The study utilized recordings of Dutch speech materials from a previously established project aimed at developing a serious game for speech disorder treatment.
- Speech Materials:
- Two three-word lists.
- Two semantically unpredictable sentences (SUS).
- Four meaningful sentences from the "Papa and Marloes" text.
- Speakers:
- Healthy Speakers: 5 native Dutch speakers (4 male, 1 female), aged 61-69.
- Dysarthric Speakers: 13 native Dutch speakers (10 male, 3 female), aged 53-75. Ten had Parkinson's disease, and three had experienced a Cerebral Vascular Accident.
- Intelligibility Ratings:
- Eleven speech therapists provided intelligibility ratings for each utterance item using a Visual Analogue Scale (VAS) ranging from 0 (not intelligible) to 100 (intelligible). Inter-listener reliability was high (>0.9).
- Acoustic-Phonetic Features:
- A total of 103 features were extracted, including 88 from the eGeMAPS set and 15 Praat features related to duration, formant frequency, pitch, and intensity.
Key Insights and Ideas
- Feature Selection is Crucial: The study highlights that not all acoustic features are equally informative. A systematic feature selection process (LASSO + stepwise regression) is essential to identify the most predictive and discriminative features, reducing noise and potential overfitting.
- Acoustic Features Link to Intelligibility: The strong correlation between actual and predicted intelligibility scores suggests that objective acoustic measurements can effectively capture subjective intelligibility perceptions.
- Specific Features Matter for Classification: While many features contribute to intelligibility prediction, specific acoustic correlates, particularly those related to F2 bandwidth and its variation, are highly effective in distinguishing between healthy and dysarthric speech.
- Interpretability Advantage: Acoustic-phonetic features offer a significant advantage over abstract ML embeddings because they are more interpretable. This allows for a better understanding of why a speaker's speech is perceived as dysarthric or unintelligible, which can inform therapeutic interventions.
- Simplicity Can Be Powerful: The finding that just two acoustic features can achieve very high classification accuracy (96.11%) is remarkable, suggesting that targeted feature extraction can lead to efficient and effective diagnostic tools.
Key Results
- Intelligibility Prediction:
- A stepwise linear regression model using 24 selected features explained 65% of the variance in intelligibility ratings (R² = 0.65).
- The correlation between actual and predicted VAS scores was strong (r = 0.80, p < .0001).
- Speech Classification:
- Using the top 17 features (ordered by significance from the regression), LDA achieved a classification accuracy of 98.06%.
- A comparable high accuracy of 96.11% was achieved using just two features: the mean of F2 bandwidth and the coefficient of variation of F2.
- The study also noted that while accuracy generally increased with more features, some features selected by stepwise regression might not always contribute to better classification.
Limitations and Future Work
- Limitations:
- Exhaustive Pairwise Testing: The exhaustive examination of every pair of features for classification might not be scalable to larger datasets or a greater number of features.
- Feature Interpretability: While some features (like F2-related ones) are interpretable, others (e.g., MFCCs) are more abstract and harder to link directly to specific speech deviations.
- Read Speech vs. Spontaneous Speech: The study used read speech, which may have different acoustic characteristics than spontaneous speech, potentially affecting generalizability.
- Data Accessibility: The dataset used was collected by others, and due to GDPR regulations, it cannot be publicly released, limiting reproducibility and further research on this specific dataset.
- Future Work:
- Larger Datasets: Replicating the analyses on larger datasets with more speakers is recommended to improve statistical power and reduce overfitting risks.
- Different Intelligibility Measures/Languages: Exploring analyses with different intelligibility measures or in other languages.
- Syntactic-Semantic Structures: Investigating features based on different syntactic-semantic structures, which might yield different regression or classification results.
- Public Datasets: Conducting future research on publicly available datasets to enhance reproducibility.
Summary updated