Uvulopalatopharyngoplasty (UPPP), an operation that enlarges the pharyngeal airway at the level of the soft palate, improves respiratory status during sleep in only 50% of patients with obstructive sleep apnea (OSA). This poor outcome suggests that narrowing of the pharyngeal airway at nonpalatal sites contributes to the obstructive process in many patients with OSA. We have used a novel endoscopic method to identify regions of the passive pharyngeal airway most susceptible to narrowing or complete closure. In order to test the hypothesis that narrowing of the passive airway at the nasopharynx predicts a favorable surgical outcome, we have preoperatively assessed the local mechanics of the passive pharyngeal airway in 18 patients with OSA undergoing UPPP. The patient population was prospectively divided into two groups: an exclusively nasopharyngeal (ENP) group, consisting of patients exhibiting narrowing only in the nasopharynx, and a not exclusively nasopharyngeal (NENP) group, consisting of patients having at least one site of narrowing outside the nasopharynx. The frequency of respiratory disturbances and arousals and the cumulative time in apnea-hypopnea were significantly reduced after surgery for the ENP group, but not for the NENP group. Improvement rate for the ENP group (86%) exceeded that for the NENP group (18%) (p < 0.01). These differences became even greater when selection criteria for the ENP group were made more restrictive (i.e., restricted to the velopharynx) or more liberal (i.e., including secondary narrowing of the oropharynx). Our results show that evaluation of passive pharyngeal mechanics identifies patients with OSA likely to improve after UPPP.

No abstract

No abstract

In concatenative speech synthesis for English, scarcity of speech data for many contexts is a serious problem. In this paper, we propose a new unit selection scheme using a decision-tree-based clustering method that combines acoustic and linguistic knowledge with statistical modeling. This approach not only allows us to find a trainable and consistent set of generalized allophonic models but also to achieve some local optimality with respect to the limited training data. To evaluate the validity of this algorithm, regression tree generation has been carried out for both vowels and consonants from 200 phonetically balanced sentences read by a female speaker. Experimental results show that regression trees offer a promising solution for the data scarcity problem.

No abstract