TL;DR
This paper investigates the timing and speed of articulator movements during plosive-vowel transitions using electromagnetic articulography (EMA). The findings reveal significant differences in gesture durations based on the articulator involved and the following vowel, offering crucial data for improving the naturalness of synthetic speech.
What Problem Does This Paper Address?
Creating natural-sounding synthetic speech, particularly in articulatory speech synthesizers, requires accurately modeling the complex movements of the vocal tract. A key challenge lies in capturing the transitional phases between sounds, known as plosive-vowel transitions. These transitions are influenced by the physical inertia of the articulators, leading to a dynamic interplay of velocity and acceleration. While acoustic signals provide some clues, precisely measuring articulator speed and gesture durations during these transitions is difficult. This paper aims to bridge this gap by providing detailed EMA measurements of these movements, which can then be used to refine control models for articulatory speech synthesizers.
Methods
The researchers employed Electromagnetic Articulography (EMA) to track the precise movements of articulators. Ten receiver coils were attached to various parts of the vocal tract, including the tongue (tip, mid, back), lips (upper and lower), and lower incisors (for jaw opening). Additional coils on the ears and nasal bone served to correct for head movements. The study focused on analyzing the "onset" and "offset" intervals of consonantal gestures within plosive-vowel transitions. The onset interval was defined as the period of increasing articulator velocity leading up to a constriction, while the offset interval represented the decreasing velocity as the constriction was released. The duration of these intervals was measured using the MVIEW software. Statistical analysis, specifically ANOVA and post-hoc Tukey-Kramer tests, was used to determine the significance of factors like articulator type ('place'), voicing, and the following vowel on gesture durations.
Datasets
The study utilized a custom dataset of 90 plosive-vowel syllables recorded from a single speaker. The syllables were embedded in the carrier sentence "Ich habe [xxx] gesagt." The plosives included [p], [b], [t], [d], [k], [g], and the vowels were [i], [e], [a], [o], [u]. Each syllable was recorded three times, and the presentation order was randomized. This dataset was specifically collected for this EMA study.
Key Insights and Ideas
- Articulator-Specific Timing: The duration of consonantal gestures is not uniform; it significantly depends on which articulator is involved (labial, apical, or dorsal). This highlights the need for articulator-specific timing parameters in speech synthesis.
- Vowel Influence on Offset: The duration of the consonant's release (offset) is influenced by the degree of opening of the following vowel. Specifically, the vowel [a], with its wide opening, led to longer offset durations.
- Bridging Acoustics and Articulation: The study demonstrates the value of EMA in providing direct measurements of articulatory movements, which are difficult to infer solely from acoustic data, especially for events like the initial stages of consonant formation.
- Data for Synthesis: The acquired data is directly intended to inform and tune the control models of articulatory speech synthesizers, aiming to improve the acoustic quality of plosive-vowel transitions.
Key Results
- Onset Durations:
- Apical plosives: 97-166ms (mean 165ms)
- Dorsal plosives: 104-197ms (mean 132ms)
- Labial plosives: 95-157ms (mean 118ms)
- The factor 'place' (articulator type) had a highly significant effect on onset durations (p<0.0001 for labial/apical and apical/dorsal differences).
- Offset Durations:
- Ranges varied from 110ms to 285ms.
- Mean offset durations: apical (215ms), dorsal (198ms), labial (184ms).
- 'Place' and 'vowel' were significant factors influencing offset durations.
- Post-hoc analysis revealed significant differences in offset durations for different vowel contexts, particularly with [a] showing longer durations.
- Articulator Velocity: Mean articulator speeds were calculated: labial (11.8 cm/s), tongue tip (12.5 cm/s), and tongue body (8.4 cm/s).
Limitations and Future Work
- Single Speaker: The study was conducted with only one subject, limiting the generalizability of the findings. Future work should include a larger and more diverse subject pool.
- Speaking Rate: The analyzed speech was at a modal speaking rate. It is unclear how speaking rate might affect articulator speed, and the authors plan to investigate this in future studies.
- Evaluation Study: The authors plan to conduct an evaluation study after implementing these findings into an articulatory speech synthesizer to assess the impact on naturalness and intelligibility. They anticipate improvements in plosive bursts and aspiration noise in CV and CVC syllables.
Summary updated