Nasal Consonant Discrimination in Infant- and Adult-Directed Speech

Bogdan Ludusan, Annett Jorschick, Reiko MazukaInterspeech 2019

TL;DR

This paper investigates whether infant-directed speech (IDS) makes nasal consonants clearer for infants compared to adult-directed speech (ADS). Using acoustic analysis and machine learning on Japanese speech, they found no significant difference in the discriminability of nasal consonants between IDS and ADS, suggesting that any acoustic enhancements in IDS might not extend to these specific speech sounds.

What Problem Does This Paper Address?

Language acquisition is a complex process, and researchers have long been interested in how the speech adults use when talking to infants (infant-directed speech, or IDS) might facilitate this process. While it's known that IDS often features exaggerated intonation, simpler vocabulary, and altered vowel acoustics (e.g., "vowel expansion"), less is understood about how consonants are treated. Specifically, this paper addresses the question of whether IDS acoustically enhances nasal consonants (/m/ and /n/) in a way that makes them easier for infants to distinguish from each other, compared to how they are produced in regular adult-to-adult speech (adult-directed speech, or ADS). Understanding these acoustic differences is crucial for building comprehensive models of language acquisition and the role of caregiver speech.

Methods

The researchers employed a two-pronged approach:

  1. Acoustic Similarity Measurement: They analyzed intervocalic bilabial (/m/) and alveolar (/n/) nasals from a Japanese corpus. For each nasal sound, they extracted nine spectral slices (feature vectors) representing the amplitude of the spectrum in the 0-2000 Hz range, which is known to contain perceptually relevant information for nasal place of articulation (nasal murmur and formant transitions). They then calculated the Pearson correlation between corresponding spectral slices of /n/ and /m/ sounds within identical vowel contexts and speech registers (IDS vs. ADS). A lower correlation would indicate greater acoustic distinctiveness.
  2. Unsupervised Machine Learning Classification: To complement the similarity analysis, they used an Expectation-Maximization (EM) algorithm to train an unsupervised classifier. This algorithm attempts to group the spectral features into two clusters (representing /n/ and /m/) without prior knowledge of the true class labels. The performance of this classifier was evaluated using the F1-score, a metric that balances precision and recall.

Datasets

The study utilized the RIKEN Mother-Infant Conversation corpus. This corpus contains:

  • Over 11 hours of infant-directed speech (IDS).
  • Approximately 3 hours of adult-directed speech (ADS).
  • Recordings of 22 Japanese mothers interacting with their 18-24 month-old infants, as well as separate recordings of mothers discussing child-rearing topics with an interviewer.
  • The data was manually transcribed and annotated at segmental and prosodic levels.

The researchers specifically selected sequences of vowel-nasal-vowel, excluding nasals at the beginning or end of prosodic phrases, and only included contexts that appeared at least five times in both IDS and ADS.

Key Insights and Ideas

  • Focus on Nasals: While vowels have been extensively studied in IDS, this paper highlights the under-researched area of consonant modifications, specifically nasal consonants.
  • Perceptually Relevant Features: The choice to focus on the lower spectral range (0-2000 Hz) for nasal murmur and formant transitions is grounded in established phonetics research, making the analysis directly relevant to perceptual distinctiveness.
  • Dual Methodology: Combining acoustic similarity measures (Pearson correlation) with machine learning classification provides a robust evaluation of whether acoustic differences translate to discriminability.
  • Within- vs. Between-Class Similarity: The study validates its similarity measure by showing higher within-class similarity (/n/-/n/, /m/-/m/) than between-class similarity (/n/-/m/), as expected.

Key Results

  • Acoustic Similarity: The Pearson correlation analysis revealed a slight decrease in similarity between /n/ and /m/ in IDS compared to ADS, but this difference was not statistically significant.
  • Machine Learning Classification: The unsupervised classifier achieved similar F1-scores for discriminating /n/ and /m/ in both IDS and ADS. A paired t-test showed that the difference in performance between registers was not statistically significant (p = 0.064).
  • Within-Category Differences: Interestingly, the study found significant differences in within-category similarity: /n/-/n/ similarity was slightly higher in IDS, while /m/-/m/ similarity was slightly lower in IDS compared to ADS. However, these opposing effects did not translate to a significant difference in the /n/-/m/ between-category similarity.

Limitations and Future Work

  • Statistical Significance: The primary limitation is that the observed acoustic differences and classification performance differences between IDS and ADS were not statistically significant for the /n/-/m/ discrimination. This suggests that, for these specific nasal consonants in Japanese, IDS does not offer a clear acoustic advantage for discrimination.
  • Within-Category Effects: The authors note that the differing trends in within-category similarity for /n/ and /m/ warrant further investigation. They hypothesize that this might be related to word-edge effects in IDS and their differential impact on /n/ and /m/.
  • Corpus Specificity: The findings are based on a specific Japanese corpus. Future work could explore these phenomena in other languages and with larger, more balanced datasets, particularly concerning the position of nasals within words.
  • Holistic Learning: The paper acknowledges that acoustic differences are only one piece of the puzzle. Future research should consider how these acoustic cues interact with other linguistic levels (e.g., prosody, lexical information) and infant preferences for IDS to form a complete picture of language learning.

Summary updated

BibTeX
@inproceedings{ludusan2019nasal,
  author = {Bogdan Ludusan and Annett Jorschick and Reiko Mazuka},
  title = {{Nasal Consonant Discrimination in Infant- and Adult-Directed Speech}},
  booktitle = {Interspeech},
  year = {2019},
}

Loading related content...

Figure