Towards interpretable speech biomarkers: exploring MFCCs.

Area of Science:

• Speech processing
• Biomedical engineering
• Clinical diagnostics

Background:

• Speech biomarkers show promise for disease detection but often lack clinical interpretability.
• Mel frequency cepstral coefficients (MFCCs) are established speech markers but are considered uninterpretable.
• Existing methods struggle to bridge the gap between complex signal processing features and clinical understanding.

Purpose of the Study:

• To investigate correlations between MFCCs and interpretable speech biomarkers.
• To enhance the clinical interpretability of MFCCs in disease detection.
• To improve the sensitivity of MFCCs to disease-induced voice alterations.

Main Methods:

• Exploration of correlations between MFCC coefficients and established speech biomarkers.
• Quantification of the MFCC2 endpoint, relating it to low- to high-frequency energy ratios.
• Analysis of MFCC2 performance across multiple datasets with adjusted computation parameters.

Main Results:

• Identified significant correlations between MFCC coefficients and interpretable speech features.
• Demonstrated that MFCC2 can be interpreted as a ratio of low- to high-frequency energy.
• Showcased increased sensitivity of MFCC2 to disease by optimizing computational parameters.

Conclusions:

• MFCCs can be made more clinically interpretable by linking them to established speech biomarkers.
• The MFCC2 endpoint offers a quantifiable measure related to disease-induced voice changes.
• Optimizing MFCC computation parameters can significantly enhance the utility of speech biomarkers for disease detection.