TL;DR
This paper explores how a simplified physical model of the human vocal tract, the Sliding Three-tube (S3T) model, can be used to generate both adult and child-like vowel sounds by adjusting its length and the fundamental frequency of the input sound. The findings highlight the crucial interplay between vocal tract length and fundamental frequency in shaping vowel perception, with implications for speech education and research.
What Problem Does This Paper Address?
Understanding how humans produce speech, particularly vowels, is a fundamental challenge in acoustics and linguistics. While sophisticated models exist, there's a need for simpler, yet effective, physical models that can be used for educational purposes and to explore the acoustic principles of speech production. Specifically, this paper addresses the gap in understanding how variations in vocal tract length (VTL) and fundamental frequency ($f_0$) contribute to the production of different vowel qualities, especially when comparing adult and child speech. Existing models often focus on adult male speech, leaving a need to explore how these parameters influence sounds perceived as childlike.
Methods
The core of this research utilizes the Sliding Three-tube (S3T) model, a simplified physical representation of the human vocal tract. This model consists of an outer and inner cylinder, where sliding the inner cylinder changes the constriction within the vocal tract, mimicking tongue movement. The researchers designed four versions of the S3T model with varying outer and inner cylinder lengths, effectively creating different overall vocal tract lengths.
To generate vowel sounds, these S3T models were coupled with a driver unit that produced input signals. Six different input signals were used: five based on impulse trains with varying fundamental frequencies ($f_0$) and durations, and one swept-sine signal for acoustic analysis. The researchers systematically varied the position of the inner cylinder (representing vocal tract configuration) and the input $f_0$ to produce a range of vowel sounds.
Acoustic analysis involved calculating spectrographic representations of the impulse responses, allowing for the visualization of spectral peaks corresponding to vocal tract resonances. These were compared against theoretical resonance curves derived from established acoustic phonetics principles.
Furthermore, a perceptual experiment was conducted where human listeners were asked to identify the vowel sound, rate their satisfaction, and guess the speaker group (adult male, adult female, child, or baby). This experiment used a subset of the recorded vowel sounds generated by the S3T models.
Datasets
The paper does not introduce new datasets in the traditional machine learning sense. Instead, it utilizes:
- Synthesized vowel sounds: Generated by the S3T models under various configurations (different lengths, different input $f_0$).
- Human listener responses: Collected from a perceptual experiment where participants identified and categorized the synthesized vowel sounds.
The synthesized sounds can be considered a novel "dataset" for exploring the acoustic properties of simplified vocal tract models.
Key Insights and Ideas
- VTL and $f_0$ are intertwined for vowel perception: The study strongly suggests that both vocal tract length and fundamental frequency are crucial for producing and perceiving different vowel qualities, especially when distinguishing between adult and child speech.
- Children's vowels are not just higher $f_0$: While higher $f_0$ is a key differentiator for childlike voices, shorter vocal tract lengths are also perceived as indicative of younger speakers.
- The S3T model is versatile: The simple S3T model, with its adjustable length and constriction, can effectively mimic a wide range of vowel sounds, making it a valuable tool for both research and education.
- Auditory system's robustness: The fact that listeners can often identify the same vowels despite variations in VTL and $f_0$ points to the human auditory system's ability to extract and separate information about vocal tract size and shape.
- Educational potential: The model's simplicity makes it ideal for hands-on science workshops, allowing children to build and experiment with their own vocal tract models.
Key Results
- The S3T models, with adjusted lengths and input $f_0$, successfully produced sounds perceived as vowels from both adult and child speakers.
- Perceptual experiments showed that as $f_0$ increased and VTL decreased, listeners were more likely to identify the speaker as a child or female.
- A strong correlation was found between the S3T model's spectral peaks and theoretical resonance curves, validating the model's acoustic behavior.
- The study identified specific combinations of VTL and $f_0$ that consistently produced a set of four vowels (/i/, /u/, /o/, /a/) with high correlation to adult male speech, demonstrating that good vowel quality can be maintained by simultaneously shortening VTL and increasing $f_0$.
Limitations and Future Work
- Model Simplification: The S3T model is a simplification of the complex human vocal tract. It does not account for all articulatory nuances, such as tongue shape, epilaryngeal constrictions, or nasalization.
- Limited Stimuli: While the study explored various combinations, the range of input signals and model configurations could be further expanded.
- Future Work: The authors express a desire to use the child-like S3T model more frequently in science workshops and classroom demonstrations. Further research could explore more detailed acoustic analyses of the recorded sounds and expand the perceptual experiments to include a wider range of participants and languages. Investigating the relationship between VTL, $f_0$, and intelligibility, particularly during voice mutation in puberty, is also suggested.
Summary updated