Relation of formants and subglottal resonances in Hungarian vowels

Tamás Gábor Csapó, Zsuzsanna Bárkányi, Tekla Etelka Gráczi, Tamás Bőhm, Steven M. LulichInterspeech 2009

TL;DR

This paper investigates the relationship between vowel formants and subglottal resonances (SGRs) in Hungarian, extending previous work on other languages. The findings suggest that SGRs act as acoustic boundaries that help define phonological distinctions in Hungarian vowels, similar to their role in English, German, and Korean. This has implications for understanding phonological theory and improving automatic speech technologies.

What Problem Does This Paper Address?

The paper explores the "Quantal Theory" of speech, which posits that certain acoustic-articulatory relationships are more stable and thus more likely to form the basis of phonological distinctions. Specifically, it focuses on the interaction between the vocal tract's resonances (formants) and the subglottal airway's resonances (SGRs). While previous studies have shown that SGRs can acoustically separate vowel categories in languages like English, German, and Korean, this relationship had not been systematically studied for Hungarian. Hungarian presents a unique challenge due to its complex vowel system, including paired long and short vowels and a series of front rounded vowels. The authors aim to determine if SGRs play a similar role in defining phonological distinctions in Hungarian as they do in other languages.

Methods

The researchers collected acoustic data from four native Hungarian speakers (two male, two female) producing nonsense words containing all Hungarian vowels. They measured the first three formants (F1, F2, F3) from the microphone signal and the first three subglottal resonances (Sg1, Sg2, Sg3) from an accelerometer attached to the neck. Formants were measured automatically and hand-corrected, while SGRs were measured manually.

To analyze the relationship between formants and SGRs, they used two primary methods:

  1. Vowel Plots: They plotted the measured formant frequencies (F1 vs. F2) for each speaker, overlaying horizontal and vertical lines representing the SGRs. This allowed for a visual inspection of how vowel clusters related to the SGR boundaries.
  2. Normalized Distributions: They normalized formant frequencies by the corresponding SGRs (e.g., F1 relative to Sg1) and plotted the distributions of these normalized values for different vowel categories (e.g., low vs. non-low vowels). This provided a quantitative measure of how well SGRs separated these categories.

They also performed Receiver Operating Characteristic (ROC) curve analysis to determine optimal frequency ranges for SGRs to separate vowel categories.

Datasets

The study utilized a custom dataset of nonsense words containing all Hungarian vowels, spoken by four native speakers. This dataset was recorded in an anechoic chamber and processed using standard acoustic analysis tools. No pre-existing benchmark datasets were used or introduced.

Key Insights and Ideas

  • SGRs as Phonological Boundaries in Hungarian: The study provides strong evidence that SGRs function as acoustic boundaries for phonological distinctions in Hungarian vowels, mirroring findings in other languages.
  • Sg2 and Front/Back Distinction: The second subglottal resonance (Sg2) consistently acts as a boundary between front and back vowels, a finding that holds across speakers.
  • Sg1 and Low/Non-Low Distinction: The first subglottal resonance (Sg1) appears to separate low vowels from non-low vowels, though this boundary is noted as being less clear, potentially due to measurement difficulties.
  • Sg3 and Front Rounded/Unrounded Distinction: The third subglottal resonance (Sg3) effectively differentiates front rounded vowels from their unrounded counterparts.
  • Vowel Space Variability: The study highlights that while SGRs provide general boundaries, individual speakers can exhibit variations in their vowel spaces, with some vowels appearing acoustically in regions not predicted by their phonological category. This suggests that while SGRs are important, other factors like co-articulation and individual speaker characteristics also play a role.

Key Results

  • The study confirmed that Sg2 reliably separates front and back vowels in Hungarian.
  • Sg1 was found to be a boundary between low and non-low vowels, although the clarity of this separation varied.
  • Sg3 was shown to differentiate front rounded vowels from front unrounded vowels.
  • ROC analysis indicated that specific frequency ranges around SGRs optimally separated vowel categories for each speaker.
  • The median SGR values for speakers generally fell within or close to the optimal ranges identified by the ROC analysis, supporting the proposed boundaries.

Limitations and Future Work

  • Measurement Challenges for Sg1: The authors acknowledge that measuring Sg1 is more difficult than Sg2 due to acoustic losses and its proximity to the fundamental frequency, which may contribute to the less clear separation of low vs. non-low vowels.
  • Speaker Variability: One speaker (F1) showed an unusually compressed vowel space, with a back vowel ([o]) appearing acoustically in the "front" region. The authors suggest further investigation is needed to determine if this is an individual anomaly or part of a broader pattern.
  • Vowel [e] Exception: The vowel [e] showed some exceptions in its relation to Sg3, which the authors attribute to its potential differences in quality and quantity compared to its long counterpart [e:].
  • Future Work: The authors suggest that their findings have implications for improving speaker normalization and other sub-problems in automatic speech recognition. Further research could explore these applications and investigate the role of SGRs in other languages with complex vowel systems.

Summary updated

BibTeX
@inproceedings{csapo2009relation,
  author = {Tamás Gábor Csapó and Zsuzsanna Bárkányi and Tekla Etelka Gráczi and Tamás Bőhm and Steven M. Lulich},
  title = {{Relation of formants and subglottal resonances in Hungarian vowels}},
  booktitle = {Interspeech},
  year = {2009},
}
Figure