Inter-dependent Categorization of Voices and Segments

Anne Cutler, Attila Andics, Zhou FangICPhS 2011

Introduction

In speech perception, a central question is whether the brain processes the linguistic message (the "what") and the speaker's identity (the "who") through independent or integrated channels. Traditional "normalization" models suggested that listeners strip away indexical voice characteristics to access an abstract phonetic code. However, contemporary research suggests these dimensions are inter-dependent. This study utilizes the Garner interference paradigm to investigate the nature of this integration, specifically focusing on how task difficulty and speaker familiarity influence the direction of interference between phonetic segments and voice identity.

Methods

The researchers employed a speeded two-alternative forced choice (2AFC) task with 30 native Dutch-speaking participants. Listeners categorized Vowel-Consonant (VC) syllables based on either a phonetic segment (e.g., identifying the vowel /ɛ/ vs. /ɒ/ or the consonant /s/ vs. /t/) or the speaker's voice (e.g., identifying "Peter" vs. "Thomas").

To test for inter-dependence, the study used the Garner manipulation:

  1. Constant Context: The irrelevant dimension remained fixed (e.g., identifying the voice while the syllable was always "ot").
  2. Varying Context: The irrelevant dimension varied across four possibilities (e.g., identifying the voice while syllables switched between "et," "es," "ot," and "os").

Notably, the study used four highly similar male speakers, requiring participants to learn and distinguish unfamiliar voices, thereby increasing the difficulty of the voice categorization task relative to the phonetic task.

Results

The experimental data revealed a clear disparity in task difficulty. Segment decisions were significantly faster and more accurate than voice decisions. As shown in the performance data, voice categorization resulted in much higher error rates (~25%) compared to segment categorization (~5%).

Mean reaction times (ms) and error rates (%) for the Garner task conditions.

The primary finding was a robust Garner effect: participants were consistently slower in the "varying" context than in the "constant" context across all conditions (p < .001). However, this interference was asymmetric. Phonetic variation disrupted voice categorization significantly more than voice variation disrupted phonetic processing.

Percentage increase in reaction times (Garner interference) for segment and voice categorization.

As illustrated in Figure 1, varying the phonetic segment increased voice reaction times by approximately 12-15%, whereas varying the voice only increased segment reaction times by about 6-8%.

Discussion

The results support the "Harder is More Vulnerable" principle. The direction of interference in speech processing is not fixed but depends on cognitive demand. Because identifying unfamiliar, similar voices is more taxing than identifying well-known phonemes, voice decisions are more susceptible to phonetic "noise."

Furthermore, the findings suggest that phonetic cues may serve as "anchors" for voice learning. When encountering a new talker, listeners likely use the phonetic shape of the speech to build a mental model of the identity. Constant phonetic variation hinders this process, making it more difficult to isolate the speaker's unique vocal signature.

Limitations and Future Work

While the study highlights the inter-dependence of voice and segment processing, several questions remain. The use of highly similar male voices was intentional but limits generalizability to more diverse acoustic environments. Future research should explore whether these asymmetries persist with familiar voices or in more complex linguistic contexts, such as continuous speech where coarticulation provides additional cues.

Summary updated

BibTeX
@inproceedings{cutler2011inter,
  author = {Anne Cutler and Attila Andics and Zhou Fang},
  title = {{Inter-dependent Categorization of Voices and Segments}},
  booktitle = {ICPhS},
  year = {2011},
}
Figure