Acoustic differentiation of l- and l-l% in switchboard and radio news speech

Heejin Kim, Tae-Jin Yoon, Jennifer Cole, Mark Hasegawa‐Johnson2006 · DOI

TL;DR

This paper investigates how speakers signal different levels of prosodic phrasing (intermediate phrase vs. intonational phrase) using acoustic cues. They found that longer nucleus duration, lower final pitch (F0), and increased creaky voice are reliable indicators of the stronger intonational phrase boundary, supporting theories of intonation that recognize multiple levels of prosodic structure.

What Problem Does This Paper Address?

Linguists and speech scientists generally agree that speech is organized into hierarchical prosodic structures, with levels like the "intonational phrase" (IP). However, there's ongoing debate about whether a finer-grained level, the "intermediate phrase" (ip), also exists and how it's acoustically signaled. While Japanese has strong evidence for this finer level (the accentual phrase), empirical evidence in English has been less clear, particularly regarding acoustic correlates beyond duration. Previous work suggested that while duration and glottalization might differentiate these levels, pitch (F0) and intensity contours were less reliable. This paper aims to provide robust acoustic evidence for distinguishing between ip and IP boundaries in English, specifically focusing on the low-toned intermediate phrase (L-) and low-toned intonational phrase (L-L%) boundaries, using both spontaneous and read speech.

Methods

The researchers analyzed speech data from two corpora: the Switchboard WS97 corpus (spontaneous telephone conversations) and the Radio News corpus (scripted broadcast speech). They focused on the phrase-final syllable rime, examining several acoustic features:

  • Nucleus Duration: The length of the vowel in the final syllable of a phrase.
  • F0 (Pitch) Measures: Beginning F0, end F0, F0 drop (end F0 - beginning F0), and F0 slope (F0 drop / duration).
  • Intensity Measures: Beginning intensity, end intensity, and intensity drop.
  • Voice Quality: Specifically, the occurrence of creaky voice (glottalization).

To ensure comparability across speakers and conditions, durations were normalized within corpora or speakers. F0 and intensity were also normalized, with a novel approach using utterance-level normalization for the Switchboard corpus to account for speaker-specific pitch ranges and paralinguistic factors. Statistical analyses (ANOVA) were used to compare these measures between L- and L-L% boundaries.

Datasets

  • Switchboard WS97 Corpus: A subset of the Switchboard telephone speech corpus, annotated with ToBI (Tones and Break Indices) labels. This is the same dataset used in previous work by the authors.
  • Radio News Corpus: Specifically, the lab news portion of the Boston University Radio News corpus, also ToBI-labeled and word-level transcribed.

Key Insights and Ideas

  • Two Levels of Phrasing are Acoustically Marked: The study provides strong evidence that speakers use distinct acoustic cues to signal different levels of prosodic phrasing (ip vs. IP).
  • F0 at Rime End is Crucial: While beginning F0 didn't reliably differentiate boundary types, the F0 value at the end of the phrase-final syllable was consistently lower for the stronger IP boundary (L-L%) compared to the weaker ip boundary (L-). This suggests a localized effect of the boundary tone.
  • Creaky Voice as a Boundary Marker: The frequency of creaky voice was significantly higher before IP boundaries (L-L%) than before ip boundaries (L-) in both corpora, reinforcing its role as a cue for stronger phrase endings.
  • Duration is a General Cue, but Degree Matters: Preboundary lengthening (longer nucleus duration) is a general cue for phrase boundaries, but the degree of lengthening was significantly greater for IP boundaries, indicating it signals boundary strength.
  • Speaker Variability Exists: While some cues (like F0 at rime end and creak) were consistent across speakers and corpora, others (like intensity contours and F0 slope/drop) showed more speaker-dependency, suggesting that speakers might utilize different combinations of cues to signal phrase boundaries.

Key Results

  • Duration: Nucleus duration was significantly longer for L-L% boundaries compared to L- boundaries in both Switchboard and Radio News corpora.
  • F0:
    • Switchboard: End F0 was significantly lower for L-L% than for L-.
  • Radio News: End F0 was lower, and F0 drop and slope were greater for L-L% than for L- for both speakers. Beginning F0 was not significantly different.
  • Intensity:
    • Switchboard: Beginning intensity was significantly lower for L-L% than for L- (for plain tokens).
  • Radio News: Speaker F2B showed significant differences in beginning and end intensity (lower for L-L%), while F1A showed no significant intensity differences for plain tokens. For creaky tokens, end intensity was significantly lower for L-L% for both speakers.
  • Creak: Frequency of creak occurrence was markedly higher for L-L% tokens than for L- tokens in both corpora.

Limitations and Future Work

  • Speaker Dependency: The authors acknowledge that the acoustic correlates of boundary strength can vary between speakers. While they analyzed speaker-dependent effects in the Radio News corpus, a similar in-depth analysis for the Switchboard corpus might reveal further nuances.
  • Higher-Level Domains: The analyses did not explicitly account for domain levels higher than the phrase (e.g., utterance boundaries). Future work could explore how these higher levels interact with phrase-level marking.
  • Generalizability: While the study used two distinct corpora (spontaneous and read), further investigation across a wider range of speaking styles and languages would be beneficial to confirm the universality of these findings.
  • Perceptual Correlates: While acoustic evidence is presented, future research could directly investigate the perceptual salience of these acoustic cues for listeners in differentiating phrase levels.

Summary updated

Abstract

Acoustic evidence for a distinction between low-toned intermediate (ip) and intonational phrase (IP) boundaries is presented from two speech corpora representing spontaneous, conversational speech and scripted broadcast speech.Robust effects of the two boundary levels are found in the phrase-final syllable rime in both corpora.Nucleus duration is longer and the F0 value at rime end is lower at IP boundaries compared to ip boundaries.Glottalization is also more frequent before an IP boundary.Other effects of boundary level on the F0 and intensity contours over the phrase-final rime are evident but variable across the two corpora.These findings support the Beckman-Pierrehumbert theory of intonation [1] in its recognition of two levels of prosodic phrasing.
BibTeX
@misc{kim2006acoustic,
  author = {Heejin Kim and Tae-Jin Yoon and Jennifer Cole and Mark Hasegawa‐Johnson},
  title = {{Acoustic differentiation of l- and l-l\% in switchboard and radio news speech}},
  year = {2006},
  doi = {10.21437/speechprosody.2006-165},
}
Figure