Acoustic pattern recognition of /s/ misarticulation by the self-organizing map

Area of Science:

• Speech processing
• Acoustic analysis
• Computational linguistics

Background:

• Psychoacoustic classification of speech sounds is crucial for understanding perception.
• Previous methods for analyzing speech acoustics lacked the ability to correlate features with subjective judgments.
• Self-organizing maps (SOMs) offer a novel approach to analyzing complex data patterns.

Purpose of the Study:

• To investigate the utility of self-organizing maps (SOMs) for analyzing acoustic features of speech.
• To correlate spectral characteristics of speech samples with psychoacoustic acceptability ratings.
• To explore the potential of SOMs for on-line visual imaging of speech data.

Main Methods:

• Utilized a self-organizing map (SOM), a neural network algorithm, trained on non-disordered speech samples.
• Analyzed fifteen-component spectral vectors derived from short-time Fast Fourier Transform (FFT) spectra at 10-ms intervals.
• Studied speech samples from 11 women, psychoacoustically classified as acceptable or unacceptable.

Main Results:

• A significant correlation was found between the location of speech samples on the SOM and their degree of audible acceptability.
• Distinct spectral features in the [s] sound samples were identified in different map locations.
• SOMs successfully depicted distinguishing spectral characteristics related to psychoacoustic judgments.

Conclusions:

• Self-organizing maps are well-suited for extracting and quantifying acoustic features that underpin psychoacoustic classifications.
• SOMs provide a valuable tool for the on-line visual representation and analysis of speech.
• This research highlights the potential of neural networks in bridging objective acoustic measurements and subjective speech perception.