Automatic intelligibility assessment of pathologic speech over the telephone

Area of Science:

• Speech-language pathology
• Biomedical engineering
• Acoustic analysis

Background:

• Objective assessment of speech intelligibility is crucial for voice and speech rehabilitation.
• Patients undergoing partial laryngectomy often experience altered speech intelligibility.

Purpose of the Study:

• To develop an objective method for assessing speech intelligibility in patients after partial laryngectomy.
• To evaluate the effectiveness of combining acoustic and linguistic features for intelligibility prediction.

Main Methods:

• 82 patients post-partial laryngectomy read standardized text, recorded via headset and telephone.
• Human intelligibility ratings were obtained from five experienced raters on a five-point scale.
• Objective evaluation used support vector regression with word accuracy (WA), word correctness (WR), and prosodic features.

Main Results:

• Word accuracy and correctness alone correlated with human ratings (|r| = 0.57–0.75).
• Combining prosodic features and WR achieved higher correlations (r = 0.79 headset, r = 0.86 telephone).
• An optimal feature subset included WR, silent pause duration, fundamental frequency standard deviation, jitter standard deviation, and voiced/total duration ratio.

Conclusions:

• Objective assessment of speech intelligibility in laryngectomy patients is feasible and can be significantly improved.
• A combination of linguistic and prosodic features provides a robust prediction of intelligibility.
• The identified feature set is effective for both high-quality (headset) and degraded (telephone) speech signals.