Evaluating acoustic representations and normalization for rhoticity classification in children with speech sound...

Area of Science:

• Speech-language pathology
• Acoustic phonetics
• Machine learning in speech science

Background:

• Accurate classification of children's speech sound production is crucial for early identification of speech sound disorders.
• Acoustic analysis offers objective measures for speech sound characterization.
• Understanding the impact of acoustic features and normalization techniques is key to developing robust speech analysis tools.

Purpose of the Study:

• To compare the effectiveness of different acoustic representations and normalization methods for classifying children's production of the rhotic sound /ɹ/.
• To evaluate the performance of formants and Mel frequency cepstral coefficients (MFCCs) in predicting rhotic versus derhotic /ɹ/.
• To assess the impact of utterance-level versus age-and-sex-based normalization on classifier accuracy.

Main Methods:

• Utilized acoustic features, including formants and MFCCs, from 350 child speakers.
• Applied z-standardization normalization, comparing within-utterance and age-and-sex-matched typical /ɹ/ data.
• Employed statistical modeling and deep neural networks for classification, with Shapley additive explanations for feature importance analysis.

Main Results:

• Age-and-sex normalization significantly enhanced classifier performance compared to utterance-level normalization.
• Clinically interpretable formants achieved performance comparable to MFCCs.
• Personalized deep neural network models reached a mean F1-score of 0.81 (SD=0.10) for participant-specific prediction.
• Shapley additive explanations identified the third formant as the most influential feature for predicting fully rhotic productions.

Conclusions:

• Age-and-sex normalization is a critical factor for improving the accuracy of automated speech analysis in children.
• Formant-based acoustic features are viable alternatives to MFCCs for deep learning applications in speech perception research.
• The findings support the use of personalized, data-driven models for accurate assessment of phonological development, with formants offering clinical interpretability.