Testing the Distinctiveness of Intonational Tunes: Evidence from Imitative Productions in American English

Eleanor Chodroff, Jennifer S. ColeInterspeech 2019

TL;DR

This paper investigates whether speakers of American English can reliably distinguish and reproduce eight distinct intonational "tunes" proposed by Pierrehumbert (1980). The study found that while speakers could differentiate between rising and non-rising contours, finer distinctions within these categories were largely lost in production, suggesting that rising/non-rising is a primary, accessible distinction in intonational grammar.

What Problem Does This Paper Address?

Understanding how intonation conveys meaning is a fundamental challenge in linguistics. While we know that variations in pitch can change the meaning of a sentence (e.g., indicating questions, statements, or emotional states), precisely how these meanings are mapped to specific intonational contours is complex. Unlike word-based language where distinct sounds map to distinct words, intonation is more fluid. Pierrehumbert's influential theory (1980) proposes a system of abstract phonological units (pitch accents, phrase accents, boundary tones) that combine to create distinct "nuclear tunes" at the end of phrases. The core question this paper addresses is whether speakers actually produce and perceive these eight theoretically distinct tunes as separate entities, or if their internal representation of intonation is more generalized.

Methods

The researchers created eight synthesized intonational contours based on Pierrehumbert's model and the TOBI (ToBI) transcription system. These synthesized contours were then played to 32 native English speakers. Participants were asked to listen to these synthesized sentences and then produce novel sentences with the same intonational pattern, under the guise of improving computer-generated speech.

Two main analyses were performed on the participants' productions:

  1. Deviation Analysis: This measured how much the participants' produced intonation contours differed from the original synthesized stimuli. This helps understand how accurately speakers could imitate the target tunes and which tunes were more difficult to reproduce.
  2. Cluster Analysis: This approach looked at the patterns in the participants' productions without direct comparison to the stimuli. It aimed to identify how many distinct groups (clusters) of intonational contours emerged naturally from the speakers' imitations, revealing their underlying perceptual and productive categories.

Datasets

The study did not use pre-existing datasets for training. Instead, they created their own synthesized stimuli.

  • Synthesized Stimuli: Eight distinct intonational contours were synthesized based on Pierrehumbert's (1980) model and TOBI approximations. These were generated using two model speakers (one male, one female) who recorded natural productions of the tunes, which were then used to derive pitch point values for the synthesis.
  • Participant Productions: 32 native English speakers produced imitations of these synthesized tunes on novel sentences. These productions formed the dataset for analysis.

Key Insights and Ideas

  • Rising vs. Non-Rising is Primary: The most striking finding is the strong evidence for a fundamental distinction between rising and non-rising intonational contours. Speakers were much better at differentiating and reproducing these broad categories than finer distinctions within them.
  • Abstract Representations May Be Generalized: The results suggest that while speakers might have abstract representations for intonation, these representations might be more generalized than the eight distinct tunes proposed by Pierrehumbert. The system seems to prioritize the broad rising/non-rising distinction.
  • Production is More Generalized Than Theory: The study highlights a potential disconnect between theoretical phonological models of intonation and actual speaker production. The theoretical model proposes fine-grained distinctions, but speakers' imitative productions tend to simplify these.
  • "Flat" Tunes are Highly Variable: The "flat" intonational contours (associated with declarative statements) showed significant variability in production, suggesting they might be more flexible or allow for greater phonetic realization options.

Key Results

  • Deviation Analysis: Participants showed significantly greater deviation from the synthesized contours for rising tunes (HHH, LHH) compared to non-rising tunes. This indicates these rising tunes were harder to imitate accurately.
  • Cluster Analysis: When analyzing all productions, the cluster analysis primarily identified two main groups: rising tunes and non-rising tunes. When the rising tunes were excluded, further analysis revealed some marginal distinctions among the non-rising tunes, but not the full eight-way distinction.
  • Specific Tune Performance: While HHH and LHH (rising tunes) were difficult to reproduce, HLL (a falling tune often used for statements) also showed high variability. LHL (a low falling tune) was the easiest to imitate accurately.

Limitations and Future Work

  • Stimulus Fidelity: The authors acknowledge that the synthesized f0 contours might not have perfectly captured the intended abstract representations, or that listeners might have struggled to access the intended meaning, hindering their ability to reproduce the target intonation.
  • Focus on Nuclear Tunes: The study focused only on the "nuclear region" of intonational phrases, excluding prenuclear pitch accents and bitonal pitch accents. A more comprehensive analysis could include these elements.
  • Imitation vs. Natural Speech: The experiment used imitation, which might differ from how speakers naturally use intonation in spontaneous speech.
  • Future Directions: The authors suggest that a follow-up experiment focusing specifically on distinctions within non-rising tunes is warranted. They also propose that the large f0 contrast in rising contours might have masked finer distinctions, and future work could explore this. The study provides a foundation for developing new hypotheses about the inventory and structure of intonational contours.

Summary updated

BibTeX
@inproceedings{chodroff2019testing,
  author = {Eleanor Chodroff and Jennifer S. Cole},
  title = {{Testing the Distinctiveness of Intonational Tunes: Evidence from Imitative Productions in American English}},
  booktitle = {Interspeech},
  year = {2019},
}
Figure