TL;DR
This paper introduces a novel acoustic analysis of word-final glottal stops in Mizo and Assam Sora, two understudied languages. By examining dynamic vocal tract characteristics (VTC) and source features, the authors reveal how glottal stop production alters vocal tract resonance and excitation, offering a new perspective beyond traditional phonetic analysis.
What Problem Does This Paper Address?
Glottal stops, characterized by a brief closure of the vocal folds, are phonemes found in many languages. While their acoustic properties have been studied, word-final glottal stops, particularly in less-researched languages like Mizo and Assam Sora, remain poorly understood. Existing research often focuses on allophonic variations or intervocalic stops. This paper aims to bridge this gap by providing a detailed acoustic characterization of word-final glottal stops in Mizo and Assam Sora, exploring how the unique articulatory demands of these stops influence both the vocal tract's resonant behavior and the glottal source's excitation characteristics.
Methods
The study employs a multi-faceted acoustic analysis:
- Dynamic Vocal Tract Characteristics (VTC): The researchers utilize a Zero Time Liftering (ZTL) based spectral estimation method to capture VTC at every sample instant. This is achieved by first multiplying the speech signal with a decaying impulse-like window function for temporal resolution. To highlight formant peaks, the Hilbert Envelope of Numerator Group Delay (HNGD) function is applied, creating an "HNGD spectrum." The Dominant Resonance Frequency (DRF) is then extracted as the strongest spectral peak corresponding to the zero crossing of the differenced smoothened HNGD spectrum. This allows for the tracking of DRF changes throughout the glottal stop and surrounding vowel.
- Source Features: To characterize the glottal source behavior, two features are extracted from the Linear Prediction (LP) residual:
- Strength of Excitation (SoE): This temporal feature quantifies the abruptness of glottal closure.
- H1-H2 Ratio: This spectral feature represents the slope of the source spectrum, indicating changes in harmonic content.
- Statistical Analysis: Mean values of SoE, H1-H2, and DRF are computed for both glottal stop (GS) and non-glottal stop (NGS) contexts to compare their distributions.
Datasets
- Mizo Speech Database: Collected from four native Mizo speakers (2 female, 2 male) at IIT Guwahati. Speakers produced 56 unique words (33 CVC, 23 CV) in a sound-attenuated booth using a Shure head-worn microphone. Annotations were done manually using Praat.
- Assam Sora Speech Database: Sourced from an existing database recorded in tea estates in Assam. This data includes recordings from fourteen Assam Sora speakers (5 female, 9 male). The recording and annotation procedures are reported to be similar to the Mizo data, ensuring comparable quality.
Both datasets were recorded at a 44.1 kHz sampling rate.
Key Insights and Ideas
- DRF as a Window into Glottal Stop Production: The study demonstrates that the Dominant Resonance Frequency (DRF) of the vocal tract changes significantly during glottal stop production. Specifically, DRF tends to increase during the closed phase and shows a more abrupt transition in the open phase compared to non-glottal stop contexts. This suggests that the increased laryngeal constriction and altered subglottal coupling during glottal stops directly impact vocal tract resonance.
- Reduced Subglottal Coupling: The research indicates a reduction in the duration of subglottal coupling during glottal stops. This is inferred from the DRF contours, where lower DRF values (associated with coupling) are observed for shorter durations in glottal stop contexts.
- Language-Specific Differences: While both Mizo and Assam Sora exhibit changes in DRF and source features during glottal stops, the salience of these changes differs. Mizo shows clearer distinctions between glottal stop and non-glottal stop contexts across all analyzed features. Assam Sora, however, shows less distinct differences, potentially due to the prevalence of creaky phonation in final vowels and less precise articulation of glottal stops in some instances.
- SoE as a Marker of Abruptness: The Strength of Excitation (SoE) is found to be higher in glottal stop contexts for Mizo, reflecting the abrupt glottal closure characteristic of these sounds. This aligns with the idea that vocal folds vibrate more irregularly and stiffly during glottal stop production.
Key Results
- Mizo: Word-final glottal stops in Mizo are characterized by a gradual increase in DRF in the glottal stop region, with a higher DRF value during the closed phase compared to the open phase. The duration of subglottal coupling is also observed to be reduced. Statistical analysis shows significant differences in SoE, H1-H2, and DRF between glottal stop and non-glottal stop contexts. A 3D plot effectively distinguishes glottal stop and non-glottal stop contexts in Mizo based on these three features.
- Assam Sora: While DRF values are generally higher in glottal stop regions, the differences are less salient compared to Mizo. SoE values show overlap between glottal stop and non-glottal stop contexts, with some non-glottal stop contexts exhibiting higher SoE due to creakiness. H1-H2 and DRF show some salience, but less so than in Mizo.
- Temporal DRF Trends: In both languages, the averaged DRF shows an increasing trend from the preceding vowel towards the glottal stop in the glottal stop context. In contrast, the non-glottal stop context in Mizo shows a decreasing DRF trend, while Assam Sora shows an increasing trend due to creaky phonation.
Limitations and Future Work
- Limitations:
- The study acknowledges that the Assam Sora data, while comparable, was from an existing database, and some glottal stops were not clearly articulated.
- The interaction between Mizo tones and glottal stops was not considered in this analysis.
- The current work focuses on word-final glottal stops and does not explore their behavior in other positions.
- Future Work:
- Investigate the complex interaction between Mizo tones and glottal stops.
- Explore automatic detection of word-final glottal stops by comparing their characteristics with other non-glottalic stop consonants in the coda position.
- Further research into the acoustic characteristics of glottal stops in other languages and positions.
Summary updated