Speech intelligibility estimation using multi-resolution spectral features for speakers undergoing cancer treatment

Area of Science:

• Speech processing
• Biomedical engineering
• Otolaryngology

Background:

• Head and neck cancer frequently impairs speech production, reducing intelligibility.
• Accurate assessment of speech intelligibility is crucial for patient rehabilitation and quality of life.

Purpose of the Study:

• To develop and evaluate a novel method for extracting spectral features to improve speech intelligibility prediction.
• To assess the effectiveness of these features in regression and classification tasks for head and neck cancer speech.

Main Methods:

• A multi-resolution sinusoidal transform scheme was employed for spectral feature extraction.
• Regression models predicted interval-scaled intelligibility scores using these features.
• Binary intelligibility classification was performed on a separate dataset.

Main Results:

• The proposed features reduced mean squared estimation error from 0.43 to 0.39 (p < 0.001) on a 1-7 scale.
• Feature inclusion improved binary intelligibility classification by 5.0 percentage points.

Conclusions:

• The novel spectral features significantly enhance the prediction of speech intelligibility.
• This method offers a promising tool for objective speech assessment in head and neck cancer patients.