Modeling speaker variation in cues to prominence using the Bayesian information criterion

Tim Mahrt, Jennifer Cole, Margaret M. Fleck, Mark Hasegawa‐Johnson2012 · DOI

TL;DR

This paper investigates how individual speakers vary in their use of acoustic cues (like duration and intensity) to signal word prominence. By applying the Bayesian Information Criterion (BIC), they found that speakers use a mix of linear and discrete cueing strategies, and importantly, that not all speakers exploit the full range of prominence distinctions available in English.

What Problem Does This Paper Address?

In spoken language, "prominence" is crucial for conveying meaning, marking important words, and structuring speech rhythmically. Phonetically, prominence is signaled by changes in duration, intensity, and pitch. While previous research has established correlations between these acoustic cues and perceived prominence, there's a gap in understanding how individual speakers differ in their use of these cues. Specifically, do speakers consistently use cues in the same way, or is there significant inter-speaker variability? This paper aims to model this variation by determining whether speakers use acoustic cues in a continuous (linear) or categorical (discrete) manner to signal prominence.

Methods

The core of this study is the application of the Bayesian Information Criterion (BIC) to model the relationship between acoustic features and perceived prominence. The authors first collected prominence judgments (p-scores) from naive English speakers on a large corpus. Then, for each acoustic feature (e.g., vowel duration, intensity), they compared two types of models:

  1. Single Gaussian Distribution: Represents a linear relationship between the acoustic cue and prominence.
  2. Two Gaussian Distributions: Represents a discrete, or piece-wise linear, relationship, suggesting a threshold where the cue's behavior changes.

The BIC was used to determine which model (single or double Gaussian) best fit the data for each feature, for each individual speaker. A positive ΔBIC score indicated that the two-distribution model was a better fit, suggesting a discrete cueing strategy at a particular p-score threshold. By analyzing these thresholds across speakers and features, the researchers could identify patterns of variation.

Datasets

  • Ohio State Buckeye Corpus: A 35,000-word subset of conversational American English speech from 27 speakers.
  • Rapid Prosody Transcription (RPT): A method used to obtain word-level prominence judgments from naive native English speakers. Groups of 15-20 speakers were asked to tag words as prominent in real-time, resulting in "p-scores" representing the proportion of listeners who marked a word as prominent.

Key Insights and Ideas

  • Speaker-Specific Prominence Strategies: The most significant finding is that speakers are not uniform in how they signal prominence. While some acoustic cues are used similarly across speakers, others show considerable individual variation.
  • Mix of Linear and Discrete Cues: For any given speaker, some acoustic features are best modeled as linearly related to prominence (continuous signaling), while others are best modeled as discrete (categorical signaling). This suggests a complex interplay of cue usage.
  • Varying "Granularity" of Prominence: The study suggests that not all speakers exploit the full range of possible prominence distinctions. Some speakers may "flatten" the prosodic hierarchy, using fewer distinct levels of prominence than are theoretically available.
  • Low p-score Threshold Dominance: Across most features, speakers tend to use a binary (two-distribution) model with a low p-score threshold. This implies a fundamental distinction between prominent and non-prominent words, with further distinctions being less universally applied.

Key Results

  • Feature-Specific Modeling: The analysis revealed that for most acoustic features, a two-distribution model (indicating discrete cueing) was preferred over a single-distribution model (linear cueing).
  • Varied Thresholds: When a two-distribution model was optimal, the p-score thresholds that best separated the two distributions varied significantly across different acoustic cues.
  • Speaker Groupings: The researchers identified distinct groups of speakers based on their cue usage patterns, with some speakers using more discrete thresholds (mid or high p-scores) in addition to the common low p-score threshold.
  • Log Duration Cues are More Linear: Log duration measures and word duration were more commonly modeled by a single distribution, suggesting a more consistently linear relationship with prominence compared to other cues.

Limitations and Future Work

  • Listener Role: The authors acknowledge that the listener plays a role in prominence perception and that further investigation is needed to understand this aspect.
  • Cognitive Representations: The study raises questions about the cognitive representations of prominence and how they are realized in speech production.
  • Production vs. Comprehension: While this study focuses on production, future work could explore how these production variations impact listener comprehension.
  • "Flattened" Hierarchy: The idea that some speakers may not fully exploit the prosodic hierarchy warrants further investigation.

Summary updated

Abstract

This study investigated speaker variation in the production of various acoustic cues of prominence, including duration and intensity measures. The Bayesian Information Criterion was used to specify a threshold distinction between cues that are linearly vs. piece-wise linearly predictors of the degree of perceived prominence. For all speakers, some features are linear and some features are discrete in the manner in which they cue prominence. However, the results also suggest that speakers differ in the number of prominence distinctions that they make. Under a metrical stress notion of hierarchically layered prominence, our result would suggest that some speakers do not exploit the full range of prominence distinctions offered in English.
BibTeX
@misc{mahrt2012modeling,
  author = {Tim Mahrt and Jennifer Cole and Margaret M. Fleck and Mark Hasegawa‐Johnson},
  title = {{Modeling speaker variation in cues to prominence using the Bayesian information criterion}},
  year = {2012},
  doi = {10.21437/speechprosody.2012-82},
}
Figure