2012 Predicting perceptual scores of speech intelligibility, articulation and voice quality for speakers with head and neck cancer

Cancer of the head and neck can have a negative effect on a speaker's speech and voice quality. In the clinical situation, a clinician-independent manner to evaluate the perceptual quality a patient's speech and voice quality is sought. To this end, models using input from automatic evaluation tools (e.g. speaker feature sets) were developed to predict perceptual scores for speech intelligibility, articulation and voice quality. The validation corpus contains longitudinal recordings of 55 speakers (before treatment, short-term and long-term follow-up) and perceptual scores completed by 13 raters.

Performance of articulation and speech intelligibility prediction models exceeds that of the average human rater; performance of voice quality prediction model is competitive. Although acoustic features selected for the articulation model overlap with features selected for the speech intelligibility model, the articulation model selects more consonant-related features and the speech intelligibility selects more vowel-related features. The voice quality model selects acoustic features related to phonation and resonance. The models show variable success in tracking trends of an individual speaker between evaluation moments.