Self-voice recognition in 4 to 5-year-old children

Sofia StrömbergssonInterspeech 2009

TL;DR

This study investigates whether 4-5 year old children can recognize their own recorded voice, finding they perform significantly above chance. Interestingly, a 1-2 week delay between recording and identification did not impact performance, but greater acoustic similarity between a child's voice and reference voices increased the likelihood of misidentification.

What Problem Does This Paper Address?

While adults are generally adept at recognizing their own recorded voices, there's a notable scarcity of research on this ability in young children. Understanding this capability is crucial, especially given the increasing use of voice recordings in speech and language therapy for children. The paper aims to fill this gap by exploring how well 4-5 year olds can identify their own voice among others and whether factors like time delay and acoustic similarity influence their performance. This has direct implications for how voice-based interventions are designed and evaluated.

Methods

The researchers developed a novel experimental setup to test self-voice recognition in young children. Participants were asked to produce recordings of 24 specific words. In a first session, they recorded their voices and then immediately participated in an identification task. In a second session, held 1-2 weeks later, they only performed the identification task.

The identification task involved presenting the child with their own recorded word alongside recordings of three other children (reference voices) and asking them to select which one was theirs. The researchers employed acoustic analysis to measure voice similarity, focusing on fundamental frequency (F0) and formant frequencies (F1-F4), as well as speaking rate. Statistical analyses, including t-tests, ANOVA, and logistic regression, were used to evaluate the impact of time delay and acoustic similarity on identification accuracy.

Datasets

The study utilized recordings of 27 children aged 4-6 years old, speaking Swedish as their native language. Three additional children (6 years old) served as reference voices. The recordings were based on a script of 24 words, all starting with /tV/ or /kV/ and having primary stress on the first syllable. No new benchmark datasets were introduced; the study generated its own data through controlled recording and identification procedures.

Key Insights and Ideas

  • Children are surprisingly good at recognizing their own voice: This challenges the assumption that such a complex auditory discrimination task might be beyond the capabilities of this age group.
  • Time is not a major factor (for this age group): The finding that a 1-2 week delay didn't affect performance is significant, suggesting that memory for one's own voice is relatively robust in this age range. This contrasts with some hypotheses about memory decay in similar tasks.
  • Acoustic similarity is a double-edged sword: While intuitively one might think greater similarity makes recognition easier, the study found that higher acoustic similarity (specifically in F0) between a child's voice and reference voices actually increased the likelihood of misidentification. This suggests that children might be more attuned to subtle differences when their own voice is very similar to others, or that the task becomes more confusing.
  • Individual variation is substantial: The wide range of performance, from chance level to near-perfect accuracy, highlights that self-voice recognition is not a uniform skill and likely depends on individual factors beyond just age.

Key Results

  • Above Chance Performance: Children performed significantly above chance level in identifying their own recorded voice. The mean correct identification rate was 78.3% on the first test and 74.6% on the second test.
  • No Significant Time Delay Effect: A paired samples t-test showed no significant difference in performance between the first test (immediate identification) and the second test (1-2 weeks later).
  • F0 Similarity Predicts Errors: Logistic regression revealed that F0 similarity was a highly significant predictor of response accuracy. An increase in the difference between a child's F0 and the most similar reference F0 increased the odds of a correct answer.
  • Spectral Similarity's Role: Spectral similarity was a significant predictor of accuracy only on the second test, suggesting a potential delayed impact or a more nuanced role compared to F0.
  • Speaking Rate Irrelevant: Similarity in speaking rate did not significantly affect children's response accuracy.

Limitations and Future Work

The authors acknowledge the large variation in performance among children, suggesting potential influences of attention, concentration, or individual aptitude. They also note that two children exhibited slightly deviant speech, which might correlate with their poorer performance, warranting further investigation into the relationship between speech production deviations and self-voice recognition.

Future work should explore developmental effects on this ability, as children of different ages might react differently to time delays. Further acoustic analysis of voice similarity could provide deeper insights into children's selection behavior. The study also suggests that while speech play might be encouraged in clinical settings, it's important to consider how it might affect the ecological validity of recordings used in therapy. The findings strongly support the use of recordings in speech and language intervention, but highlight the need for careful consideration of individual differences and acoustic properties.

Summary updated

BibTeX
@inproceedings{strombergsson2009self,
  author = {Sofia Strömbergsson},
  title = {{Self-voice recognition in 4 to 5-year-old children}},
  booktitle = {Interspeech},
  year = {2009},
}
Figure