Analysis of Glottal Stop in Assam Sora Language

Sishir Kalita, Luke Horo, Priyankoo Sarmah, S.r. Mahadeva Prasanna, S. DandapatInterspeech 2016

TL;DR

This paper introduces a novel set of voice source features to characterize glottal stops in Assam Sora, a low-resource language. By analyzing acoustic signals, the authors demonstrate that these features can effectively distinguish glottal stops from adjacent vowel sounds, paving the way for improved speech processing in languages with limited data.

What Problem Does This Paper Address?

Glottal stops, produced by the abrupt closure of vocal folds, are phonetically significant but challenging to analyze acoustically, especially in continuous speech. While spectrograms offer limited insight due to the lack of supralaryngeal articulator movement, understanding the laryngeal voice source is crucial. This is particularly true for low-resource languages like Assam Sora, where existing acoustic models are scarce, and the phonetic realization of glottal stops can be variable. The paper aims to bridge this gap by developing and evaluating robust voice source features to accurately characterize intervocalic glottal stops in Assam Sora, which have been observed to have diverse production patterns.

Methods

The study employs a range of voice source features extracted from the estimated voice source signal, derived using Iterative Adaptive Inverse Filtering (IAIF). Key techniques include:

  • Temporal Domain Features:
    • Quasi-Open Quotient (QOQ): Measures the relative open time of the glottis.
  • Normalized Amplitude Quotient (NAQ): Characterizes the glottal closing phase, reflecting phonation and vocal intensity.
  • Spectral Domain Features:
    • H1-H2 Ratio: Describes the spectral slope, indicating voice tension.
  • Harmonic Richness Factor (HRF): Quantifies the harmonic content in the voice source spectrum.
  • Novel and Advanced Features:
    • Maxima Dispersion Quotient (MDQ): A wavelet-based feature designed to capture the sharpness of glottal closure.
  • Normalized Cross-Correlation Coefficient (NCC): Used to quantify period-to-period irregularity in glottal cycles.
  • Zero Frequency Filter (ZFF) evidence: A feature derived from a ZFF signal to de-emphasize vocal tract characteristics and highlight laryngeal dynamics.
  • Normalized Jitter and Logarithm Peak Normalized Excitation Strength (LPNES): Features known for their robustness in characterizing glottal stops, particularly LPNES derived from ZFF output.

A robust Glottal Closure Instant (GCI) detection method, SEDREAMS (SE-VQ), was used to handle non-modal phonation effectively.

Datasets

The study utilized a custom dataset of 14 Assam Sora disyllabic words with intervocalic glottal stops between identical vowels. These words were spoken by 12 native Assam Sora speakers (male and female, aged 23-40) and recorded in isolation and within a sentence frame. The recordings were made at a sampling frequency of 44.1 kHz, 24 bits, in WAV format, and later manually annotated using Praat.

Key Insights and Ideas

  • Variability in Glottal Stop Production: Contrary to expectations of a simple pitch dip, intervocalic glottal stops in Assam Sora can exhibit either a pitch dip or a pitch peak, highlighting the need for more sophisticated analysis beyond basic pitch and amplitude contours.
  • Salient Voice Source Features: The study demonstrates that a combination of temporal and spectral voice source features, along with novel features like MDQ and ZFF evidence, are highly effective in distinguishing glottal stops from adjacent vowels.
  • Sharp Glottal Closure: Features like NAQ and MDQ suggest a sharper and more abrupt glottal closure during glottal stops compared to vowels, which is a key characteristic.
  • Irregularity in Glottal Cycles: The NCC feature confirms period-to-period irregularities in glottal cycles during glottal stop production, potentially due to irregular vocal fold vibration.
  • Statistical Significance: A Mann-Whitney U test confirmed the statistical significance of the differences observed between glottal stops and their adjacent vowels across multiple features.

Key Results

  • Discrimination Power: The analyzed voice source features, including QOQ, NAQ, H1-H2 ratio, HRF, NCC, MDQ, ZFF evidence, LPNES, and jitter, showed significant discrimination between glottal stop and vowel regions.
  • Feature Variability: Features like QOQ, HRF, and NCC exhibited higher variability in the glottal stop region compared to vowel regions, suggesting that this variability itself can be a useful feature for glottal stop detection.
  • LPNES and Jitter: LPNES showed higher variation in glottal stop regions, and normalized jitter values were significantly higher for glottal stops (mean 1.081) than for vowels (mean 0.36), with high variance for glottal stops (1.43 vs. 0.09 for vowels).
  • Statistical Significance: A Mann-Whitney U test confirmed a significant difference (< 0.005) between glottal stop and adjacent vowel regions for all tested features.

Limitations and Future Work

  • Limited Dataset Size: The study acknowledges the limitation of a small dataset due to the low-resource nature of Assam Sora, which restricts the development of comprehensive acoustic models.
  • Focus on Intervocalic Stops: The analysis primarily focused on intervocalic glottal stops, excluding complete glottal stops due to their limited occurrence in the dataset.
  • Future Directions: The authors suggest that the developed features can be used for automatic detection and classification of glottal stops in continuous speech. Further research could explore these features in other low-resource languages and investigate their application in tasks like speech synthesis and recognition. The potential for using feature variability as an additional detection cue is also highlighted.

Summary updated

BibTeX
@inproceedings{kalita2016analysis,
  author = {Sishir Kalita and Luke Horo and Priyankoo Sarmah and S.R. Mahadeva Prasanna and S. Dandapat},
  title = {{Analysis of Glottal Stop in Assam Sora Language}},
  booktitle = {Interspeech},
  year = {2016},
}

Loading related content...

Figure