TL;DR
This paper investigates how the complex interplay between the vocal tract and the glottal excitation source influences the perception of "trill" sounds. By synthesizing speech with manipulated source and system characteristics, they found that the glottal excitation source plays a more significant role than the vocal tract shape in how we perceive these sounds.
What Problem Does This Paper Address?
Producing speech sounds involves two main components: the vocal tract, which shapes the sound, and the glottal excitation source, which is the vibration of the vocal folds. For most speech sounds, the vocal tract's resonances are considered dominant in shaping the perceived sound. However, certain sounds, like "trills" (e.g., the rolled 'r' in Spanish or Italian), involve rapid, dynamic changes in the vocal tract, specifically the tongue tip vibrating against the roof of the mouth. This rapid movement is hypothesized to influence the glottal excitation source itself. The authors aim to disentangle the contributions of these two components to the perception of trill sounds, addressing the question of whether the vocal tract's dynamic changes or the resulting changes in glottal pulsing are more critical for our auditory experience.
Methods
The researchers employed a speech synthesis approach to isolate the effects of the vocal tract system and the glottal excitation source.
- Feature Extraction:
- Vocal Tract System: Linear Predictive Coding (LPC) was used to capture the dynamic changes in the vocal tract shape. This is a standard technique in speech analysis.
- Glottal Excitation Source: Zero-Frequency Filtering (ZFF) was used to extract features related to the glottal excitation, specifically the "strength of excitation" (SoE) and the "pitch period" (To) at each glottal closure (epoch). The SoE is related to the slope of the ZFF signal around an epoch.
- Speech Synthesis Scenarios: Four scenarios were created for synthesizing trill sounds:
- Source and System Retained: Both original vocal tract and glottal source characteristics were used.
- Source Only Retained: The original glottal source characteristics were used, but the vocal tract system was altered.
- System Only Retained: The original vocal tract characteristics were used, but the glottal source was altered.
- Source and System Changed: Both vocal tract and glottal source characteristics were altered.
- Source Modification: To create the "changed" source, the authors averaged the pitch period and SoE around each epoch. This smoothed out the rapid variations characteristic of trills, making the excitation sequence more regular and less dynamic.
- System Modification: For the "changed" system, LPCs computed over a longer frame (100ms) were used, effectively averaging out the rapid vocal tract changes that occur during a trill cycle.
Datasets
The study did not use pre-existing large-scale speech databases. Instead, they recorded custom speech data:
- A sustained trill-approximant pair.
- Trill sounds produced at four different places of articulation (bilabial, dental, alveolar, post-alveolar).
This data was recorded with the help of an expert male phonetician. The synthesized speech files were then used for perceptual evaluation.
Key Insights and Ideas
- Glottal Source Dominance: The most striking finding is that the glottal excitation source appears to be more perceptually significant than the vocal tract system for trill sounds. When the glottal source was preserved, listeners rated the synthesized speech as more similar to the original, even if the vocal tract was altered.
- System-Source Coupling: While the glottal source is more dominant, the study emphasizes that neither component alone is sufficient. The interaction and coupling between the vocal tract dynamics and the glottal excitation are crucial for producing and perceiving trill sounds.
- Trill vs. Approximant Perception: The authors suggest that the dynamic changes in the glottal source are more pronounced in trills than in approximants, which helps distinguish them perceptually.
- Place of Articulation Matters (Subtly): While the glottal source was generally more important, the study noted that dental and alveolar trills were perceived as closer to their natural counterparts when both source and system were retained, suggesting some influence of vocal tract configuration on perception for specific trill types.
Key Results
The perceptual evaluation, conducted with 20 speech researchers, yielded quantitative similarity scores:
- Experiment 1 (Trill-Approximant Pair):
- Retaining both Source and System (Scenario 1): Average similarity score of 3.95.
- Retaining Source Only (Scenario 2): Average similarity score of 3.48.
- Retaining System Only (Scenario 3): Average similarity score of 2.82.
- Changing both Source and System (Scenario 4): Average similarity score of 1.75.
- Experiment 2 (Four Places of Articulation):
- Retaining both Source and System (Scenario 1): Average similarity score of 3.35 across all four trill types.
- Retaining Source Only (Scenario 2): Average similarity score of 2.79.
- Retaining System Only (Scenario 3): Average similarity score of 2.33.
- Changing both Source and System (Scenario 4): Average similarity score of 1.33.
These scores consistently show that preserving the glottal source information leads to higher similarity ratings compared to preserving only the vocal tract system.
Limitations and Future Work
- Limited Subject Pool: The perceptual evaluation was conducted with a small group of 20 subjects, all of whom were speech researchers. This might introduce bias, and results may differ with a broader, non-expert listener group.
- Synthetic Data: While the synthesis approach is powerful for isolating variables, the naturalness of the synthesized speech might not perfectly replicate real-world speech.
- Future Work: The authors suggest their findings can be applied to automatic trill spotting, trill-based language discrimination, and a deeper understanding of the production and perception of various apical trill sounds. Further research could explore the specific acoustic features of the glottal source that are most critical for trill perception.
Summary updated