Extrinsic normalization for vocal tracts depends on the signal, not on attention

Matthias Sjerps, James M. McQueen, Holger MittererInterspeech 2012

TL;DR

This paper investigates why listeners normalize for speaker-specific vocal tract characteristics more effectively with speech than with non-speech sounds. They found that increasing listener attention to non-speech precursors did not enhance normalization effects, suggesting that the signal's inherent acoustic properties, rather than attentional focus, are the primary drivers of this phenomenon.

What Problem Does This Paper Address?

When we listen to speech, our brains are remarkably adept at adjusting for variations in how different people speak – their unique vocal tract characteristics. This "extrinsic normalization" allows us to understand speech consistently, even when speakers have very different voices. Researchers have observed this effect with speech, but it's often weaker or absent when using non-speech sounds that mimic speech's acoustic properties. This discrepancy raises a question: why the difference? One hypothesis was that listeners simply don't pay as much attention to non-speech sounds as they do to speech. This paper directly tests this hypothesis by manipulating listener attention to non-speech stimuli to see if it can boost normalization effects.

Methods

The researchers replicated two experiments from a previous study ([5]) that used spectrally rotated speech targets and speech targets with manipulated precursors. The key innovation in this study was the introduction of an "attentional task." Participants were instructed to refrain from responding to the target sound if a specific dip in amplitude occurred in the precursor sound. This "catch trial" served to direct participants' attention to the precursor. The normalization effect was measured by observing shifts in vowel categorization boundaries. They used linear mixed-effects regression models to analyze the data, comparing the effects of precursor type (high vs. low F1), stimulus step, and the presence or absence of the attentional task.

Datasets

The study utilized synthesized speech stimuli based on recordings of Dutch vowels. Specifically, they created a six-step continuum between /pit/ and /pet/ based on variations in the first formant (F1). Precursor sounds were derived from a Dutch sentence and manipulated to have either high or low F1 contours. For Experiment 1, these precursors were spectrally rotated. For Experiment 2, they were not spectrally rotated but underwent further manipulations (flat pitch, removal of low-amplitude parts, temporal reversal of syllables). The study also directly compared their results to data from a previous study ([5]) that used similar stimuli but without the attentional manipulation.

Key Insights and Ideas

  • Attention is not the key: The most significant finding is that increasing listener attention to non-speech precursors did not lead to stronger normalization effects. This directly challenges the hypothesis that a lack of attention is the sole reason for weaker normalization with non-speech sounds.
  • Bottom-up signal properties dominate: The results strongly suggest that the inherent acoustic characteristics of the signal itself are the primary determinants of normalization strength, rather than top-down cognitive factors like attention.
  • Experience with speech matters: The authors propose that listeners develop a specialized ability to normalize for speech due to extensive experience with its spectrotemporal regularities. This learned sensitivity allows for more robust compensation.
  • "Speechiness" isn't just subjective: While signals need to be acoustically similar to speech to elicit normalization, this similarity is based on objective, bottom-up signal properties, not on a listener's subjective impression of "speechiness."

Key Results

  • Experiment 1 (Spectrally Rotated Speech): Normalization effects were observed, and the strength of these effects was comparable to the previous study ([5]) where no attentional task was present. The attentional manipulation did not significantly alter the normalization effect.
  • Experiment 2 (Manipulated Speech-like Sounds): No significant normalization effect was found in this experiment, even with the attentional task. While this aligns with some findings in [5], the difference between their current results and [5] was not statistically significant, suggesting that attention did not worsen the effect either.

Limitations and Future Work

The authors acknowledge that while they have rejected the attention hypothesis, the exact "low-level acoustic properties" that drive stronger normalization with speech remain to be fully elucidated. Future work could focus on dissecting these specific acoustic features and exploring how perceptual learning mechanisms might contribute to this specialized sensitivity to speech signals. They also note that while the attentional manipulation didn't increase normalization, it didn't necessarily decrease it either, leaving room for nuanced interactions.

Summary updated

BibTeX
@inproceedings{sjerps2012extrinsic,
  author = {Matthias Sjerps and James M. McQueen and Holger Mitterer},
  title = {{Extrinsic normalization for vocal tracts depends on the signal, not on attention}},
  booktitle = {Interspeech},
  year = {2012},
}
Figure