Consonant-Vowel Transition Models Based on Deep Learning for Objective Evaluation of Articulation.

Area of Science:

• Speech processing
• Acoustic phonetics
• Computational linguistics

Background:

• Spectro-temporal dynamics in consonant-vowel (CV) transitions offer reliable articulation cues.
• Existing methods for assessing articulation precision may lack objectivity or broad applicability.

Purpose of the Study:

• To propose an objective measure of precise articulation (OAM) by analyzing CV transitions.
• To validate the OAM against perceptual assessments in various speech contexts.

Main Methods:

• Analyzing CV transition regions around vowel onsets.
• Utilizing a convolutional neural network (CNN) pre-trained for consonant classification.
• Deriving the OAM from CNN posteriors.

Main Results:

• The OAM demonstrated a correlation with perceptual measures of articulation.
• This correlation was observed in adult dysarthric speech.
• The OAM was also validated in children with cleft lip/palate and accented English speech.

Conclusions:

• The proposed objective articulation measure (OAM) provides a robust and objective assessment of speech articulation.
• The OAM's effectiveness is confirmed across diverse speech populations, including clinical and non-native speakers.