Listeners’ Ability to Identify the Gender of Preadolescent Children in Different Linguistic Contexts

Shawn Nissen, Sharalee Blunck, Anita Dromey, Christopher DromeyInterspeech 2019

TL;DR

This study reveals that listeners can accurately identify the gender of preadolescent children from their speech, with accuracy increasing significantly as the linguistic context expands from individual sounds to full discourse. This suggests that while acoustic cues are present, suprasegmental features and potentially learned sociophonetic patterns play a crucial role in gender perception in young children.

What Problem Does This Paper Address?

The ability of adults to distinguish between male and female speakers is well-established, but the underlying mechanisms, especially for younger children, remain less understood. While anatomical differences in vocal tracts are known to contribute to adult gender differences, preadolescent children exhibit minimal such dimorphism. This raises the question: what cues do listeners use to identify the gender of children whose vocal anatomy is still developing? This paper investigates how different levels of linguistic context—from isolated phonemes to conversational discourse—influence listeners' ability to accurately perceive a child's gender, aiming to shed light on the relative importance of acoustic-phonetic versus suprasegmental speech features.

Methods

The researchers collected speech samples from 10 monolingual American English-speaking children (5 male, 5 female) aged between 8 and 9 years old. These samples were categorized into four linguistic contexts:

  • Segments: Isolated voiceless fricatives (/s/, /f/) and four corner vowels (/i/, /æ/, /u/, /a/).
  • Words: Three monosyllabic and three bisyllabic words.
  • Sentences: Six pre-defined sentences read by the children.
  • Discourse: Approximately 1-minute segments of conversation about academic preferences.

A group of 20 adult listeners (12 female, 8 male) with no history of speech or language disorders evaluated these stimuli. Listeners were asked to identify the speaker's gender and rate their confidence on a 1-10 scale. The study employed a repeated measures ANOVA to analyze the data, using arcsine transformations for percentage data to meet statistical assumptions. Mean speaking fundamental frequency (F0) was also calculated for each child.

Datasets

The study utilized a custom dataset of speech recordings from 10 preadolescent children. This dataset was structured into four distinct linguistic contexts (segments, words, sentences, discourse) for experimental manipulation. No new benchmark dataset was introduced; rather, existing speech recording and analysis methodologies were applied.

Key Insights and Ideas

  • Context is King: The most striking finding is the dramatic increase in gender identification accuracy as linguistic context expands. This strongly suggests that listeners rely on more than just the inherent acoustic properties of individual sounds.
  • Beyond Anatomy: Given the limited anatomical sex differences in preadolescent vocal tracts, the high accuracy points towards learned sociophonetic factors and suprasegmental cues (like intonation, rhythm, and speaking rate) playing a significant role.
  • Vowels vs. Fricatives: Listeners were more accurate in identifying gender from vowels than from fricatives, even though both were identified well above chance. This might relate to the more stable acoustic correlates of vowels (formant frequencies) compared to the more variable spectral characteristics of fricatives.
  • F0 Not the Sole Determinant: While there were differences in mean F0 between male and female children, the correlation between a speaker's F0 and listener accuracy was non-significant. This implies that pitch alone is not the primary driver of gender identification in this age group.
  • Listener Gender Matters (for Females): Female listeners were less accurate at identifying the gender of female speakers compared to male speakers, whereas male listeners showed no such bias. This could hint at subtle differences in how genders perceive gender-specific speech cues.

Key Results

  • Overall gender identification accuracy across all contexts was high, averaging 90.66%.
  • Accuracy increased significantly with linguistic context: 83.42% for segments, 90.00% for words, 93.00% for sentences, and 99.00% for discourse.
  • Individual speaker identification accuracy ranged from 82.37% to 98.68%.
  • Vowel segments yielded higher accuracy (87.13%) than fricative segments (76.00%).
  • Significant differences were found in fricative identification between /s/ and /f/, but not among the four vowels.
  • Mean speaking F0 for females was 241.89 Hz (SD=20.87) and for males was 231.30 Hz (SD=10.14).

Limitations and Future Work

The authors acknowledge that while the study provides valuable insights, the sample size of children (n=10) and listeners (n=20) is relatively small. Future research could benefit from larger, more diverse participant pools across a wider age range. The study also suggests exploring the specific impact of individual acoustic parameters (e.g., duration, spectral mean, intensity) on gender perception, potentially through synthetic speech manipulation. Furthermore, a more detailed spectral and acoustic analysis of speaker productions, coupled with perception studies, could illuminate the intricate links between production and perception of gender-related speech cues. The authors also note that the efficacy of the arcsine transformation is still under evaluation.

Summary updated

BibTeX
@inproceedings{nissen2019listeners,
  author = {Shawn Nissen and Sharalee Blunck and Anita Dromey and Christopher Dromey},
  title = {{Listeners’ Ability to Identify the Gender of Preadolescent Children in Different Linguistic Contexts}},
  booktitle = {Interspeech},
  year = {2019},
}
Figure