A harmonic-cancellation-based model to predict speech intelligibility against a harmonic masker

Area of Science:

• Acoustic Phonetics
• Auditory Perception
• Speech Processing

Background:

• Harmonic maskers, unlike noise, possess structure allowing potential speech intelligibility release.
• Mechanisms like spectral glimpsing and harmonic cancellation are proposed for fundamental frequency (F0) segregation but lack empirical validation.
• The relative contributions of these mechanisms to predicting behavioral data remain unexplored.

Purpose of the Study:

• To develop and validate a computational model predicting speech intelligibility in the presence of harmonic maskers.
• To investigate the roles of spectral glimpsing and harmonic cancellation in F0 segregation and their predictive power for human listeners.
• To optimize model parameters using existing experimental data on speech reception thresholds.

Main Methods:

• A novel speech intelligibility model incorporating both spectral glimpsing and harmonic cancellation was developed.
• The model was applied to fit data from two experiments measuring speech reception thresholds with varying harmonic masker F0 and harmonicity.
• Key model parameters were optimized by maximizing the correlation between model predictions and the experimental data.

Main Results:

• The developed model accurately predicted the influence of masker F0 and harmonicity on speech intelligibility.
• Model predictions showed a high correlation (0.99) with the behavioral data across both experiments.
• Mean and maximum absolute prediction errors were notably low, below 0.5 dB and 1 dB, respectively.

Conclusions:

• The model successfully captures the complex interplay of spectral glimpsing and harmonic cancellation in predicting speech intelligibility against harmonic maskers.
• The findings support the proposed mechanisms and demonstrate their utility in quantitatively explaining auditory perception of speech in complex spectral environments.
• This validated model offers a powerful tool for understanding and predicting speech intelligibility in scenarios involving spectrally rich maskers.