Objective child vocal development measurement with naturalistic daylong audio recording

Dongxin Xu, Jill Gilkerson, Jeffery A. RichardsInterspeech 2012

TL;DR

This paper introduces an objective, automated method for measuring child vocal development using daylong audio recordings from natural home environments. By analyzing features like phonetic patterns and prosody, the system can estimate a child's "vocalization age" with high accuracy and even identify potential developmental disorders like autism.

What Problem Does This Paper Address?

Traditionally, assessing child vocal development has relied on subjective observations and clinical evaluations. This approach can be time-consuming, prone to bias, and may not capture the full spectrum of a child's vocal behavior in their natural environment. The authors aim to bridge this gap by developing an objective, unobtrusive, and scalable method that leverages daylong audio recordings to extract rich developmental information. This objective measurement is crucial for early identification of developmental disorders, monitoring progress, and informing clinical practice and parenting.

Methods

The core of this research lies in the automated analysis of daylong audio recordings. Here's a breakdown of their approach:

  • Data Collection: Lightweight digital recorders were worn by children for extended periods (typically 9-16 hours) in their natural home settings.
  • Sound Segmentation: Signal processing and pattern recognition techniques were used to automatically segment the audio into different sound categories: key-child (the child wearing the recorder), adult, noise, and silence.
  • Feature Extraction: A variety of features were extracted from the "key-child" segments, including:
  • Phonetic Development Features: Analysis of phone-level sound categories (e.g., consonant-like, vowel-like, non-speech-like sounds) and specific phonetic sounds like nasals ([M], [N], [NG]) and voiced stops ([B], [D], [G]).
  • Unsupervised Sound Categories: Using K-means clustering to identify emergent sound categories without pre-defined labels.
  • Sequence Features: Analyzing patterns in sequences of sounds.
  • Prosodic Features: Examining characteristics like duration and loudness (dB) of sounds.
  • Developmental Trend and Group Difference Analysis: Statistical methods (correlation with age, Welch's two-sample t-tests) were employed to identify how these features change with age and how they differ between groups of children (typically developing, language delay, and autism).
  • Vocalization Age Estimation: A linear regression model was built using the extracted features to predict a child's chronological age, termed "vocalization age."
  • Autism Identification: An Adaboost method (an approximation to logistic regression) was used to classify children as typically developing or having autism based on the extracted features.

Datasets

The study utilized a dataset of 1363 naturalistic daylong audio recordings from 106 typically developing (TD) children, 49 children with language delay (LD), and 71 children with autism (ASD). The age ranges for these groups were:

  • TD: 8-48 months (mostly >12 months)
  • LD: 10-44 months (mostly 14-41 months)
  • ASD: 16-48 months (mostly 25-48 months)

The paper notes that recordings containing therapy time were removed from the ASD group to avoid confounding effects.

Key Insights and Ideas

  • Objective Measurement is Feasible: The study demonstrates that objective, automated measurement of child vocal development is not only possible but also highly informative, moving beyond subjective assessments.
  • Rich Developmental Information in Daylong Audio: Even seemingly simple daylong audio recordings contain a wealth of data that can reveal subtle developmental trends and group differences.
  • Phonetic Development as a Key Indicator: The analysis of specific phonetic sounds (nasals, voiced stops) and general sound categories (consonant-like, vowel-like) showed clear age-related progression and significant differences between developmental groups, particularly highlighting slower development in ASD children.
  • Vocalization Age as a Novel Metric: The concept of "vocalization age" provides a new, objective metric for assessing a child's vocal development trajectory, independent of chronological age.
  • Potential for Early Disorder Identification: The high accuracy achieved in identifying autism suggests that these objective vocal features can serve as valuable biomarkers for early detection of developmental disorders.

Key Results

  • Vocalization Age Estimation: A correlation of 0.84 was achieved between the estimated vocalization age and the chronological age for typically developing children using leave-one-child-out cross-validation. For LD and ASD children, the vocalization age was consistently underestimated, indicating developmental delays.
  • Autism Identification: The system achieved 94% accuracy at an equal-error rate point for identifying childhood autism using the Adaboost method.
  • Developmental Trends: Features like the percentage of consonant-like sounds showed a significant positive correlation with age across all groups, with TD children developing most rapidly.
  • Group Differences: Significant differences were observed in various features between TD, LD, and ASD children, often reflecting slower or atypical development in the latter two groups. For example, ASD children produced longer non-speech-like sounds and had higher dB levels for vowel-like sounds.

Limitations and Future Work

  • Interpretation of Mathematical Features: While features derived from mathematical approaches like Principal Component Analysis (Figure 6) showed utility, their direct physical meaning can be challenging to interpret. The authors suggest that tracing down actual related physical meanings is possible.
  • Spectrum Features: Other spectrum features (spectrum tilt, spectrum entropy) were studied but not fully reported in this paper, indicating potential for future analysis.
  • Broader Application: The authors believe this approach has great potential for child development research, clinical practice, and parenting, suggesting future work could focus on refining these applications.
  • Further Exploration of Features: The paper mentions that a large number of features were studied, but only a sample is presented. Further investigation into other features could yield additional insights.

Summary updated

BibTeX
@inproceedings{xu2012objective,
  author = {Dongxin Xu and Jill Gilkerson and Jeffery A. Richards},
  title = {{Objective child vocal development measurement with naturalistic daylong audio recording}},
  booktitle = {Interspeech},
  year = {2012},
}
Figure