Low dimensional measurement of vowels using machine perception.

Area of Science:

• Acoustics
• Computational Linguistics
• Machine Learning

Background:

• Traditional sound analysis often lacks linguistic nuance.
• Dimensionality reduction techniques can simplify complex acoustic data.
• Neural networks excel at feature extraction from spectrograms.

Purpose of the Study:

• To develop a method for generating linguistically motivated, low-dimensional sound measurements.
• To integrate the feature extraction capabilities of Convolutional Neural Networks (CNNs) with model-based dimensionality reduction.
• To create sound measurements that reflect perceptual characteristics.

Main Methods:

• Training a CNN to predict linguistic sound categories from spectrograms.
• Defining idealized category models as probability distributions in a low-dimensional space.
• Matching CNN output probabilities to idealized models to determine sound measurements.

Main Results:

• Demonstrated the method using monophthongal vowel categories, producing two-dimensional measurements.
• The resulting CNN's perceptual characteristics were found to be similar to human listeners.
• Successfully generated low-dimensional measurements from complex sound data.

Conclusions:

• The proposed method effectively combines deep learning with model-based approaches for sound analysis.
• This technique produces linguistically relevant and perceptually aligned low-dimensional sound representations.
• The findings suggest potential applications in speech processing and auditory perception research.