Automatic evaluation of tracheoesophageal substitute voices

More about the book

In 20 to 40 percent of laryngeal cancer cases, total laryngectomy necessitates the removal of the entire larynx, resulting in the loss of the natural voice and primary means of communication. A common voice restoration method involves a shunt valve, or "voice prosthesis," creating a tracheoesophageal (TE) substitute voice. Evaluating this substitute voice is crucial for therapy progress, but current assessments are subjective and reliant on the therapist's experience. This thesis explores the use of automatic methods to provide an objective evaluation of substitute voices. While established measures focus on sustained vowels, this research extends to analyzing entire words and sentences, essential for assessing intelligibility in real communication scenarios. Automatic word recognition methods were employed on standard texts read by test subjects, comparing word recognition rates to human expert evaluations. A prosody module was utilized to extract acoustic information and individual speaking characteristics. The findings revealed that the correlation between human and automatic ratings was comparable to the agreement among human raters. Enhancements to automatic recognition were noted with distant-talking recordings using mu-law features and modified Mel-Frequency Cepstrum Coefficients (MFCC). Additionally, using artificially reverberated training data could improve recognition rates, paving the way for therapy sessions witho