Perception on pitch reset at discourse boundaries

Hsin-Yi Lin, Janice FonInterspeech 2010

TL;DR

This study investigates how listeners perceive and utilize pitch resets to identify boundaries in speech. The findings suggest that listeners categorize pitch resets into broad groups rather than relying on precise measurements, and that the amount of reset, not its absolute pitch height, is the crucial factor for boundary detection.

What Problem Does This Paper Address?

Speech is not a continuous stream of sounds; it's segmented into meaningful units. Identifying these boundaries is crucial for understanding spoken language. Previous research has established that speakers use prosodic cues, like pitch changes, to signal these boundaries. Specifically, a "pitch reset" – where the pitch contour rises after a period of decline – has been observed to correlate with discourse boundaries. However, it was unclear how listeners actually perceive and use this pitch reset cue to detect boundaries. Do they rely on the exact amount of the pitch rise, or is there a more general perception at play? This paper aims to bridge this gap by experimentally testing how listeners' reaction times to boundary detection are affected by varying degrees of pitch reset.

Methods

The researchers designed an experiment using synthesized speech stimuli. They created 18-syllable utterances from nonsense syllables ("bu," "di," "ga") to avoid semantic influences. Boundaries were placed at the 7th, 9th, or 11th syllable. Crucially, they manipulated the "pitch reset" at these boundaries, varying the amount of the pitch rise after the preceding declining contour. They also varied the "pitch register" (overall pitch level) and the "sentence length" before the boundary.

The core of the experiment involved presenting these synthesized stimuli to participants and asking them to press a button as soon as they heard the start of the second sentence (i.e., after the boundary). Their reaction times (RT) were measured from the moment of the pitch reset. A three-way repeated ANOVA was used to analyze the effects of pitch reset, pitch register, and sentence length on these reaction times.

Datasets

Not discussed in the paper. The stimuli were synthesized by the researchers.

Key Insights and Ideas

  • Categorical Perception of Pitch Reset: Listeners don't seem to perceive pitch resets on a fine-grained scale. Instead, they group them into broad categories (e.g., small, medium, large resets). This suggests that listeners don't need to precisely measure the pitch rise to detect a boundary.
  • Amount Over Absolute Height: The study strongly suggests that the amount of pitch reset is what matters for boundary detection, not the absolute pitch height after the reset. This aligns with the idea that speakers manipulate the change in pitch to signal structure.
  • Pitch Declination as a Pre-Signal: The observation of negative reaction times (RTs faster than the pitch reset itself) hints that the preceding pitch declination might also play a role, potentially by allowing listeners to anticipate upcoming boundaries as the declination slows down.
  • Subordinate Role of Pitch Register: While pitch register did have an effect, it was generally subordinate to other cues and disappeared when other information (like sentence length) was more prominent. This further supports the idea that the relative change (reset) is more important than the absolute pitch level.

Key Results

  • Main Effects: All three manipulated factors (Pitch Reset, Pitch Register, Sentence Length) had significant main effects on reaction times.
  • Pitch Reset: Larger pitch resets generally led to shorter reaction times, but this effect became less pronounced with very large resets (above 20 Hz). This supports the categorical perception idea.
  • Sentence Length: Longer sentences before a boundary consistently resulted in shorter reaction times, suggesting listeners had more time to prepare for the boundary.
  • Interactions: Significant interactions were found between:
  • Register x Reset: The effect of pitch register was only significant for very small resets.
  • Sentence Length x Register: The effect of register was only significant for the shortest sentence length (7 syllables).
  • Sentence Length x Reset: The effect of pitch reset was more pronounced for shorter sentences and disappeared for the longest sentences (11 syllables).
  • Categorical Groupings: Post-hoc tests revealed that listeners' responses fell into roughly three categories of pitch reset: 10 Hz, 15-20 Hz, and 25 Hz and above.

Limitations and Future Work

  • Artificial Stimuli: The use of synthesized nonsense syllables, while controlled, might not fully reflect the complexity of natural speech.
  • Limited Scope of Cues: The study focused primarily on pitch reset. Future work could investigate the interplay of pitch reset with other prosodic cues like duration and amplitude, as well as non-prosodic cues like semantics and syntax.
  • Pitch Declination's Role: The paper acknowledges the potential role of pitch declination in anticipating boundaries and suggests further investigation into how pitch declination, in addition to pitch reset, affects boundary perception and discourse hierarchies.
  • Listener Variability: While the study excluded some subjects, further exploration into individual differences in prosodic perception could be valuable.

Summary updated

BibTeX
@inproceedings{lin2010perception,
  author = {Hsin-Yi Lin and Janice Fon},
  title = {{Perception on pitch reset at discourse boundaries}},
  booktitle = {Interspeech},
  year = {2010},
}
Figure