Daily auditory environments in French-speaking infants: A longitudinal dataset

Estelle Hervé, Clément François, Laurent PrevotACL 2024

TL;DR

This paper introduces a novel longitudinal dataset of daylong audio recordings from French-speaking infants, collected using the LENA system. By comparing LENA's automated metrics with other analysis tools and replicating previous findings, the study validates the reliability of these tools for studying early language development in French-speaking contexts and provides a rich resource for future research.

What Problem Does This Paper Address?

Understanding the "daily auditory environment" of infants is crucial for studying language development. While many studies have explored this using daylong recordings, most have focused on English or Spanish-speaking families. Furthermore, validating the automated analysis tools used for these recordings, particularly the LENA system, across different languages and contexts is essential. This paper addresses these gaps by: 1) creating and describing a new, longitudinal French-speaking infant dataset, 2) evaluating the reliability of LENA's metrics against other computational tools and human annotations (where possible), and 3) replicating established findings on infant language development trajectories using this new dataset.

Methods

The study utilizes the Language ENvironment Analysis (LENA) system for data collection. This system involves a wearable digital language processor (DLP) that records up to 16 hours of audio per day. The raw audio is then processed by the LENA software to automatically segment and label sound events into categories like target child (CHN), adult male (MAN), adult female (FAN), etc., providing metrics such as Adult Word Count (AWC) and Child Vocalization Count (CVC).

To validate LENA's output, the researchers employed two additional computational tools:

  • Voice-Type Classifier (VTC) from the ChildProject framework, which uses speech diarization to identify sound activity segments and map them to categories similar to LENA's.
  • Vocalization Maturity Analysis (VCM), which further refines VTC output by assessing the acoustic features of speech to classify it as canonical, non-canonical, cries, or other sounds.

The study also involved comparing LENA's metrics with those derived from VTC and VCM, and replicating established findings from previous research (e.g., Warlaumont et al., 2014) to assess the reliability and utility of the collected data and analysis pipeline.

Datasets

The primary dataset introduced in this paper consists of daylong audio recordings from 20 French-speaking infants, collected longitudinally from 3 to 24 months of age. This dataset comprises over 8286 hours of recordings. The data collection was conducted in collaboration with three daycare centers in southeastern France.

The paper also references and utilizes findings from existing datasets and studies, particularly those involving English, Spanish, Mandarin, Korean, and Dutch-speaking populations, to compare their results and validate their methodology.

Key Insights and Ideas

  • LENA's reliability extends to French: The study demonstrates that LENA's automated metrics are reliable for analyzing auditory environments in French-speaking infants, showing strong correlations with other computational tools.
  • Age is a primary driver of language development: The research confirms that infants' speech production significantly increases with age, a finding consistent across different analysis tools and robust even when controlling for demographic factors like gender, socioeconomic status, and linguistic environment.
  • The "social feedback loop" is present in French: The study replicates the finding that caregiver responsiveness to infant vocalizations plays a crucial role in shaping the child's subsequent speech production, reinforcing the importance of interactional dynamics.
  • Computational tools offer a viable alternative to manual annotation: Given the ethical and practical challenges of manual annotation for large-scale daylong recordings, the study highlights the utility of computational tools like LENA, VTC, and VCM for extracting meaningful linguistic and interactional data.

Key Results

  • Child speech production increases with age: The target child's speaking time ratio, as measured by both LENA and VTC, showed a significant positive correlation with age.
  • VCM confirms speech, not just noise, increases: The VCM analysis revealed that the increase in child vocalizations with age was specifically due to an increase in canonical and non-canonical speech, rather than just cries or noise.
  • Temporal contingencies are age-dependent: The study found significant positive relationships between age and temporal contingencies (e.g., CHILD>ADULT and ANY>CHILD), indicating that interactions become more structured as infants develop.
  • Adult responses influence child speech: Replicating previous work, the study confirmed that adult responses to child speech-related utterances tend to elicit more speech-related follow-ups from the child, supporting the "social feedback loop" hypothesis.
  • High correlation between LENA and VTC: The study reported strong correlations (e.g., r = 0.862 for child speaking time ratio) between LENA and VTC metrics across various categories, underscoring the consistency of these automated analysis approaches.

Limitations and Future Work

  • Absence of manual annotation: A primary limitation acknowledged is the lack of manual annotation for the raw audio data, which was due to legal and ethical constraints. This constraint was mitigated by cross-validating with other computational tools and replicating existing literature.
  • Data sharing: While the code for analysis is made available, the raw audio data itself cannot be shared due to privacy concerns. However, the derived metrics will be made available in the LLDC public repository.
  • Future linguistic characterization: The authors plan to extend their analysis to more nuanced linguistic characterizations, including phonological measures and content-based metrics, leveraging improvements in automatic speech recognition for daylong recordings.
  • Cross-analyses with experimental data: The paper mentions that other experimental data were collected longitudinally on the same infants, opening up possibilities for cross-analyses between the auditory environment characterization and other developmental aspects.

Summary updated

BibTeX
@inproceedings{herve2024daily,
  author = {Estelle Hervé and Clément François and Laurent Prevot},
  title = {{Daily auditory environments in French-speaking infants: A longitudinal dataset}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2024},
}
Figure