Acoustic Correlates of Phonation Type in Chichimec

Anneliese Kelterer, Barbara SchupplerInterspeech 2019

TL;DR

This paper presents the first acoustic investigation of phonation type contrasts (modal, breathy, creaky) in the Oto-Manguean language Chichimec. It reveals that Cepstral Peak Prominence (CPP) is a robust measure for distinguishing these types across vowels and speakers, while spectral tilt measures show more complex interactions with vowel height and speaker sex.

What Problem Does This Paper Address?

Many languages exhibit phonological contrasts in voice quality, known as phonation types. While modal voice is universal, breathy and creaky voice are less common and can be phonemically contrastive. Understanding the acoustic correlates of these phonation types is crucial for phonetics and phonology. However, most acoustic studies have focused on low vowels, leaving the behavior of high vowels under-explored. Furthermore, the influence of speaker sex and individual variation on these acoustic measures is not fully understood. This paper addresses these gaps by investigating the three-way phonation contrast in Chichimec, specifically including high vowels and analyzing differences between male and female speakers. It also examines how phonation type manifests over the temporal course of a vowel, which is known to be important in some Oto-Manguean languages.

Methods

The study employed acoustic analysis of recorded speech from four Chichimec speakers (two female, two male). The primary acoustic measures used were:

  • Spectral Tilt: Quantified by two measures:
  • H1-H2: The difference in amplitude between the first and second harmonics.
  • H1-A1: The difference in amplitude between the first harmonic and the first formant. These measures are expected to reflect laryngeal constriction, with steeper slopes (higher values) for breathy voice and flatter slopes (lower values) for creaky voice compared to modal voice.

  • Cepstral Peak Prominence (CPP): A measure of the harmonics-to-noise ratio, reflecting the regularity of vocal fold vibration. Lower CPP values are expected for breathy and creaky voice due to increased noise or irregular vibration.

The recorded vowels were segmented into four equal portions to analyze temporal changes. Linear mixed-effects regression models were used to statistically analyze the data, considering phonation type, tone, vowel portion, speaker sex, and vowel height as factors. To manage complex interactions, separate models were built for each combination of speaker sex and vowel type (male/female, /a/, /i/).

Datasets

The study utilized a newly recorded dataset of 30 Chichimec nouns spoken by four native speakers (2 female, 2 male). These nouns were elicited within a carrier phrase. The dataset included approximately 441 target vowels, with individual vowel portions analyzed. No pre-existing benchmark datasets were used for evaluation.

Key Insights and Ideas

  • CPP is King (for Chichimec): Cepstral Peak Prominence (CPP) emerged as the most reliable acoustic measure for distinguishing between modal, breathy, and creaky vowels in Chichimec, performing consistently across both sexes and vowel types.
  • Temporal Dynamics Matter: Phonation type in Chichimec is not static throughout the vowel. Non-modal phonation (breathy and creaky) is predominantly realized in the latter half of the vowel, providing acoustic evidence for previous linguistic observations.
  • Spectral Tilt is Tricky: Spectral tilt measures (H1-H2, H1-A1) showed more complex and variable results. While they could distinguish phonation types, their effectiveness was influenced by vowel height and speaker sex.
  • High Vowels and Women's Voices are Different: The acoustic correlates of phonation in high vowels (/i/) for female speakers were particularly unexpected, sometimes showing reversed tendencies compared to predictions based on laryngeal constriction. This suggests that spectral tilt measures might not be universally applicable to high vowels, especially in women's speech.
  • Sex Differences in Acoustic Correlates: While CPP was robust, spectral tilt measures showed some sex-specific patterns, with women's high vowels presenting a particular challenge for interpretation.

Key Results

  • CPP: Consistently distinguished modal from non-modal vowels across all speakers and vowel types, particularly in the latter half of the vowel. While it didn't differentiate breathy from creaky, its overall robustness was a key finding.
  • Spectral Tilt (H1-H2, H1-A1):
    • For women's low vowels (/a/), spectral tilt measures were better indicators than for men.
  • For men's high vowels (/i/), spectral tilt could distinguish breathy from modal, but not creaky from modal.
  • For women's high vowels (/i/), spectral tilt measures showed unexpected tendencies, with breathy /i/ sometimes having lower values than modal /i/, and creaky /i/ having higher values.
  • For men's low vowels (/a/), spectral tilt measures were generally not effective in distinguishing phonation types.

Limitations and Future Work

  • Speaker Variability: While the study controlled for sex, the authors note that individual differences among speakers (e.g., age, idiosyncratic speech patterns) might explain some of the variability, particularly in men's low vowels. Future research could benefit from differentiating between individual speakers more granularly.
  • High Vowel Interpretation: The unexpected results for women's high vowels suggest that spectral tilt measures may not be universally suitable for characterizing phonation in these contexts. Further research is needed to fully understand the interaction of voice source and vocal tract in these cases.
  • Limited Measures: The study focused on specific spectral tilt measures. Future work could explore a wider range of acoustic measures to gain a more comprehensive understanding.
  • Temporal Dynamics: While the study incorporated temporal analysis, further investigation into the precise timing and acoustic realization of non-modal phonation across different languages and speakers is warranted.

Summary updated

BibTeX
@inproceedings{kelterer2019acoustic,
  author = {Anneliese Kelterer and Barbara Schuppler},
  title = {{Acoustic Correlates of Phonation Type in Chichimec}},
  booktitle = {Interspeech},
  year = {2019},
}
Figure