TL;DR
This study investigates the long-held assumption that vocal tract length is a primary driver of individual differences in speech sound characteristics (formants and pitch). Using MRI data from fifteen Japanese male speakers, the researchers found surprisingly weak correlations between vocal tract length and body height, and even weaker correlations with most formant frequencies. The findings suggest that other factors, not just vocal tract length, are crucial for shaping unique voice characteristics.
What Problem Does This Paper Address?
For decades, speech science has operated under the intuitive notion that a speaker's physical size, particularly their vocal tract length, directly influences the acoustic properties of their voice. This is evident in the clear differences observed between adult males, females, and children. However, the precise nature and strength of these correlations within a homogenous group, like adult males with already developed vocal tracts, remained unclear. Specifically, this paper aims to quantify the relationship between vocal tract length, body height, and key acoustic features (formant frequencies and pitch) for the five Japanese vowels, seeking to understand what truly drives individual variations in voice.
Methods
The researchers utilized a magnetic resonance imaging (MRI) database of fifteen adult Japanese male speakers.
- Vocal Tract Length Extraction: A novel method, proposed by Takemoto et al. [4], was employed to define vocal tract length. This involved manually identifying the vocal fold line on midsagittal MRI images, computing a contour map of the vocal tract region, and then calculating the spline curve passing through the centroids of these contours to represent the vocal tract midline. This provided a precise measurement of the glottis-to-lips length.
- Acoustic Feature Extraction: Due to loud MRI scanning noise, speech sounds recorded prior to the scan were used.
- Formant Frequencies: The first four formant frequencies (F1-F4) were extracted from the log spectral envelopes of vowel segments using unbiased log spectral estimation.
- Pitch Frequency: The Pitch Contour function of WaveSurfer was used for pitch extraction.
- Correlation Analysis: Standard statistical methods were used to calculate correlation coefficients (r) between vocal tract length, body height, age, and the extracted acoustic features.
Datasets
The study used a pre-existing database: the "ATR vocal tract MRI data for Japanese vowels." This database consists of midsagittal MRI scans of fifteen adult male Japanese speakers producing the five Japanese vowels (/a/, /e/, /i/, /o/, /u/), along with their corresponding speech recordings and self-reported body height.
Key Insights and Ideas
- Body Height vs. Vocal Tract Length: The intuitive link between body height and vocal tract length, often assumed to be strong, is surprisingly weak in adult males. This suggests that while overall body size might influence vocal tract development, it's not a direct predictor within a developed adult population.
- Vocal Tract Length and Formants (The /e/ Exception): While vocal tract length showed little correlation with formant frequencies for most vowels, a notable exception was found for the vowel /e/. For this vowel, lower formant frequencies (F1 and F2) exhibited negative correlations with vocal tract length. This is attributed to the relatively simple, uniform tube-like shape of the vocal tract during /e/ production, making it more sensitive to length variations.
- Pitch and Body Height: A weak negative correlation was observed between pitch frequency and body height. This hints that vocal fold length might track body height to some extent, even in adults.
- Age as a Factor: The study suggests that age might play a role in vocal tract length variability in adults, potentially due to laryngeal descent in older individuals.
Key Results
- Vocal Tract Length and Body Height: The correlation coefficient between mean vocal tract length and body height was a mere r = 0.08.
- Vocal Tract Length and Age: A weak positive correlation was found between vocal tract length and age (r = 0.51).
- Vocal Tract Length and Formant Frequencies: For most vowels, correlations between vocal tract length and the first four formant frequencies were weak. However, for the vowel /e/, F1 and F2 showed negative correlations with vocal tract length (F1: r = -0.65, F2: r = -0.56).
- Pitch Frequency and Body Height: A weak negative correlation was found between pitch frequency and body height (r = -0.61).
Limitations and Future Work
- Sample Size and Homogeneity: The study acknowledges that the sample size of fifteen adult males is small. Furthermore, the findings are specific to this group and may not generalize to females or younger populations.
- Future Work: The authors suggest reinforcing these findings with a larger number of speakers and extending the analysis to adult female speakers to investigate gender-specific correlations. They also propose further research into the interplay of physical and acoustic factors for a more comprehensive understanding of voice characteristics.
Summary updated