Identification of pathological voices using glottal noise measures

Area of Science:

• Speech science
• Acoustic phonetics
• Biomedical engineering

Background:

• Vocal noise quantification is crucial for diagnosing voice disorders.
• Existing methods, including fundamental frequency (F0)-dependent and independent measures, have limitations.

Purpose of the Study:

• To compare the efficacy of F0-dependent and F0-independent vocal noise measures.
• To identify the most accurate method for classifying normal versus pathological voice samples.

Main Methods:

• Investigated four F0-dependent (time-domain and spectral) and two F0-independent measures.
• Utilized linear prediction (LP) modeling for F0-independent measures.
• Tested on a database of sustained vowel samples from 53 normal and 175 pathological talkers.

Main Results:

• LP-model-based measures significantly outperformed other methods.
• A parameter quantifying spectral flatness of the unmodeled component achieved 96.5% classification accuracy.

Conclusions:

• Linear prediction modeling offers a highly effective approach for vocal noise quantification.
• Spectral flatness analysis of LP residual is a promising biomarker for voice pathology detection.