Phone duration modeling for speaker age estimation in children

Area of Science:

• Speech processing
• Developmental linguistics
• Machine learning

Background:

• Automatic speaker age estimation is crucial for personalized content and interactive experiences.
• Estimating children's age from speech is challenging due to limited data and high variability.
• Existing methods often adapt adult speech techniques, which may not be optimal for children.

Purpose of the Study:

• To propose a novel technique for automatic speaker age estimation in children.
• To exploit temporal variability, specifically phone durations, as a biomarker for children's age.
• To develop regression models for predicting age in children from kindergarten to grade 10.

Main Methods:

• Utilizing phone duration distributions derived from forced-alignment of speech and transcripts.
• Training regression models to predict speaker age based on these phone duration features.
• Evaluating the robustness and portability of the proposed features on two children's speech datasets.

Main Results:

• Phone durations were found to contain significant developmental information relevant to children's age.
• The proposed features demonstrated robustness and portability across different speech signal conditions.
• Analysis identified specific phonemes most influential in estimating children's speaker age.

Conclusions:

• Phone duration analysis is a promising approach for automatic speaker age estimation in children.
• The developed features are effective even in low-data scenarios, addressing a key challenge in child speech research.
• This technique offers a valuable tool for applications requiring age-appropriate speech interaction with children.