Robust Estimation of Hypernasality in Dysarthria with Acoustic Model Likelihood Features

Area of Science:

• Speech-language pathology
• Acoustic phonetics
• Computational linguistics

Background:

• Hypernasality is a common symptom in motor-speech disorders, affecting both voiced and unvoiced sounds.
• Acoustic manifestations of hypernasality are highly variable, challenging clinical and automated estimation.
• Existing methods using engineered features or machine learning on limited datasets have limitations.

Purpose of the Study:

• To develop novel acoustic features for improved hypernasality estimation.
• To capture complementary dimensions of hypernasality in speech.
• To validate the generalizability of these features across different speech disorders.

Main Methods:

• Developed two acoustic models trained on a large corpus of healthy speech.
• Model 1: Measures nasal resonance in voiced sounds.
• Model 2: Measures articulatory imprecision in unvoiced sounds.

Main Results:

• The proposed acoustic features demonstrated specificity to hypernasal speech.
• Features generalized across different dysarthria corpora (e.g., Parkinson's to Huntington's disease).
• Features also generalized from neurologically disordered speech to cleft palate speech.

Conclusions:

• The novel acoustic features offer a robust method for hypernasality estimation.
• These features show excellent generalizability across various speech disorders.
• This approach holds potential for improved diagnosis and monitoring of motor-speech disorders.