Introduction
Hypernasality—the perception of excessive nasal resonance during speech—is a critical clinical marker for various motor-speech disorders (dysarthria), including Parkinson’s disease (PD), Amyotrophic Lateral Sclerosis (ALS), and Huntington’s disease (HD). It results from velopharyngeal dysfunction (VPD), where the soft palate fails to adequately seal the nasal cavity during oral speech sounds.
Despite its clinical importance, the current "gold standard" for assessment remains clinician perception. This approach is inherently subjective, expensive, and prone to high inter-rater variability. Automated acoustic analysis offers a potential solution, but traditional methods often struggle with two main issues: overly simplistic engineered features that fail to capture complex acoustic signatures, and supervised machine learning models that overfit to specific disease-related speech patterns rather than isolating the symptom of hypernasality itself.
Methodology: The NAP Framework
To address these challenges, we propose a novel framework called Nasalization Articulation Precision (NAP). This approach utilizes a hybrid architecture that combines domain-specific acoustic modeling with minimally-supervised learning.
As shown in Figure 1, the system is divided into three primary stages:
- Preprocessing: Audio is processed using forced alignment to map specific phonemes to the acoustic signal. This requires a known transcript of the spoken text.
- Acoustic Model Learning: Instead of training on pathological speech, we train two distinct models on the large-scale LibriSpeech corpus of healthy English speech.
- The Nasalization Model: Uses Gaussian Mixture Models (GMMs) to learn the distribution of acoustic patterns for voiced phonemes in "oral" and "nasal" states, calculating a log-likelihood ratio to detect abnormal nasalization in vowels.
- The Articulation Model: Employs a Hidden Markov Model (HMM-GMM) to measure the "Goodness of Pronunciation" for unvoiced consonants (e.g., /p/, /t/). In hypernasal speech, these sounds often lose precision as air pressure leaks through the nasal cavity.
- Supervised Learning: The features extracted from these healthy-speech-derived models are fed into a simple, interpretable linear ridge regression model trained on a dysarthric corpus with clinical labels.
Experimental Results
The NAP-Linear model was evaluated against several baselines, including deep neural networks (PASE-NN) and traditional MFCC-based models. We utilized a diverse dysarthric corpus (75 speakers) and a control group of speakers with Cleft Lip and Palate (CLP) to ensure the model isolated hypernasality from general speech degradation.
As detailed in Figure 5, the NAP-Linear model consistently outperformed all automated baselines. In a "leave-one-speaker-out" cross-validation, the model achieved a Pearson Correlation Coefficient (PCC) of 0.722, matching the performance of human clinicians. Crucially, "leave-one-disease-out" tests demonstrated that the model generalizes across different neurological conditions (e.g., training on PD and ALS to test on HD), confirming that it captures the specific acoustic signature of hypernasality rather than disease-specific noise.
Discussion and Future Work
The NAP framework represents a significant step toward objective, robust clinical assessment of hypernasality. By using healthy speech as a baseline, the model avoids the overfitting common in small pathological datasets.
A current limitation is the reliance on forced alignment, which requires a known transcript and can fail in cases of extremely severe dysarthria. Future research will focus on developing transcript-free models to enable remote, spontaneous speech monitoring via mobile devices, further increasing the accessibility of these diagnostic tools.
Summary updated