F0 and the Perception of Prominence

Tim Mahrt, Jennifer Cole, Margaret M. Fleck, Mark Hasegawa-JohnsonInterspeech 2012 · DOI

Abstract

This study investigates the role of fundamental frequency (F0) in the perception of word prominence in spontaneous speech. By comparing statistical models using the Bayesian Information Criterion (BIC), the researchers demonstrate that while duration functions as a gradient "slider," F0 and intensity act as discrete "switches" for identifying prominent words.

Introduction

In linguistics, "prominence" describes how certain words are acoustically highlighted to signal importance, new information, or contrast. Despite its significance, the reliability of F0 as a cue in spontaneous speech and the fundamental nature of prominence—whether it is a binary "on/off" feature or a multi-layered gradient—remain debated. This research addresses these conflicts by analyzing a large corpus of spontaneous speech to determine which model best fits human perception data.

Methodology

The study utilized a 35,000-word subset of the Ohio State Buckeye Corpus, featuring natural American English conversations. The analytical pipeline involved:

  1. Acoustic Extraction: Raw, log-transformed, and locally normalized F0 measures were extracted from words, stressed vowels, and final vowels.
  2. Rapid Prosody Transcription (RPT): 15–20 naive listeners marked prominent words in real-time to generate a p-score (0 to 1) for each word, representing the proportion of listeners who perceived it as prominent.
  3. Statistical Modeling: The Bayesian Information Criterion (BIC) was used to compare Single-Gaussian (gradient) and Two-Gaussian (discrete) models to determine the best fit for the data.

Results and Insights

The Importance of Normalization and Anchor Points

A key finding is that raw F0 values are poor predictors of prominence due to natural sentence declination. Local normalization—comparing a word's pitch to its immediate neighbors in a 3- or 5-word window—is essential for F0 to be a meaningful cue. Furthermore, F0 measurements taken from the stressed vowel were twice as predictive as those taken from the final vowel.

Table 1: Correlation (r^2) of various F0 features with perceived prominence. Stressed vowel features with local normalization show the highest predictive power.

Discrete vs. Gradient Cues

The most significant insight is the "dual nature" of acoustic cues. BIC analysis favored the two-Gaussian model for F0 and Intensity, suggesting they function as categorical triggers or "switches." In contrast, duration is gradient; as the duration of a word increases, its perceived prominence increases linearly, acting more like a "slider."

Table 2: Threshold classifications for various acoustic features, indicating discrete (Low/High) vs. gradient (None) behavior.

Normalized log F0 of the stressed vowel achieved $r^2$ values up to 0.056. While seemingly small, this is substantial for single-feature correlations in unscripted speech. The optimal "split" for F0 occurred at a low p-score, indicating its strength in distinguishing non-prominent words from the rest of the discourse.

Conclusion

The findings suggest that prominence serves both phonological metrical stress (gradient) and pragmatic focus (discrete). While F0 and intensity signal that a word is prominent, duration may signal the degree of that prominence. Future work will explore the interaction between these "switches" and "sliders" across different types of focus, such as contrastive versus new information.

Summary updated

BibTeX
@inproceedings{mahrt2012f,
  author = {Tim Mahrt and Jennifer Cole and Margaret M. Fleck and Mark Hasegawa-Johnson},
  title = {{F0 and the Perception of Prominence}},
  booktitle = {Interspeech},
  year = {2012},
  doi = {10.21437/interspeech.2012-634},
}

Loading related content...

Figure