Computing low-dimensional representations of speech from socio-auditory structures for phonetic analyses.

Area of Science:

• Linguistics
• Speech Science
• Machine Learning

Background:

• Low-dimensional representations (e.g., formants, spectral moments) are crucial for phonetic and phonological analysis.
• Existing methods often rely on traditional acoustic features.

Purpose of the Study:

• To present a novel framework for computing low-dimensional speech representations based on manifold learning.
• To demonstrate the framework's application in analyzing children's sibilant fricative production.

Main Methods:

• Utilizing manifold learning to map high-dimensional speech data to low-dimensional spaces.
• Employing the phoneigen package for constructing manifolds and data mapping.
• Comparing manifold-derived representations with standard acoustic features.

Main Results:

• Demonstrated the process of constructing manifolds and mapping speech data.
• Showcased how manifold structure influences learned low-dimensional representations.
• Found manifold-based representations effective in phonetic analysis, particularly for sibilant fricatives.

Conclusions:

• The proposed framework offers a powerful approach to speech representation learning.
• Manifold learning captures complex linguistic and socio-indexical information in speech.
• This method holds potential for advancing speech analysis and modeling.