Related Experiment Videos
Statistical analysis of word-initial voiceless obstruents: preliminary data
K Forrest1, G Weismer, P Milenkovic
1Speech Motor Control Laboratories, Waisman Center, University of Wisconsin, Madison.
The Journal of the Acoustical Society of America
|July 1, 1988
Summary
This study developed a statistical method to classify word-initial voiceless obstruents. The procedure accurately identified voiceless stops and sibilants across different speakers and genders.
Area of Science:
- Speech Acoustics
- Phonetics
- Statistical Pattern Recognition
Background:
- Accurate classification of speech sounds is crucial for understanding speech production and developing speech technologies.
- Voiceless obstruents present unique acoustic challenges due to their brief duration and complex spectral characteristics.
Purpose of the Study:
- To describe and evaluate a statistical procedure for classifying word-initial voiceless obstruents.
- To assess the effectiveness of acoustic features derived from spectral analysis for phonetic classification.
- To determine if classification models generalize across genders.
Main Methods:
- Acoustic analysis using Fast Fourier Transforms (FFTs) on monosyllabic words produced by ten speakers.
- Calculation of spectral moments (mean, variance, skewness, kurtosis) from linear and Bark-transformed spectra.
- Application of discriminant analysis to classify voiceless stops, fricatives, and sibilants.
Main Results:
- A statistical model achieved 92% accuracy in classifying voiceless stops, with high cross-gender generalization (94%).
- Classification of voiceless fricatives was below 80% accuracy.
- Classification of voiceless sibilants reached 98% accuracy using Bark-transformed spectra, also showing cross-gender validity.
Conclusions:
- Statistical analysis of acoustic moments provides an effective method for classifying certain voiceless obstruents, particularly stops and sibilants.
- The developed models demonstrate robustness across genders, suggesting generalizability in speech acoustics.
- Spectral moments, especially from Bark-transformed spectra, are valuable features for phonetic classification tasks.