Characterization of inter-speaker articulatory variability: A two-level multi-speaker modelling approach based on MRI...

Area of Science:

• Linguistics
• Speech Science
• Biomedical Engineering

Background:

• Speech communication requires shared articulatory and acoustic codes.
• Inter-individual differences in vocal tract morphology and articulation strategies create variability.
• Characterizing and modeling speaker-independent articulatory strategies is a challenge.

Purpose of the Study:

• To develop a multi-speaker modeling approach for speaker-independent articulatory strategies.
• To investigate the relationship between vocal tract morphology and articulatory strategies.
• To quantify inter-speaker articulatory variability.

Main Methods:

• A two-level modeling approach: statistically-based linear articulatory models controlled by a low-dimensionality speaker model.
• Utilizing inter-speaker correlations between morphology and strategy.
• Manual segmentation of vocal tract articulator contours from midsagittal MRI data of 11 French speakers uttering 62 vowels and consonants.

Main Results:

• Multi-speaker models explained 66%-69% of variance with root-mean-square errors of 0.36-0.38 cm in a leave-one-out procedure.
• Inter-speaker variability is primarily linked to vocal tract morphology rather than idiosyncratic strategies.
• Articulatory components adapt to individual vocal tract morphology.

Conclusions:

• The proposed multi-speaker modeling approach effectively captures articulatory variability.
• Vocal tract morphology is a dominant factor in inter-speaker speech differences.
• Understanding these relationships can improve speech synthesis and analysis.