Nehodí se? Vůbec nevadí! U nás můžete do 30 dní vrátit
S dárkovým poukazem nešlápnete vedle. Obdarovaný si za dárkový poukaz může vybrat cokoliv z naší nabídky.
30 dní na vrácení zboží
In conversational dialogue applications it is§critical to understand the requests accurately.§However, the performance of current speech§recognition systems are far from perfect. In order to§function effectively with imperfect speech§recognition, an accurate confidence scoring mechanism§should be employed. To determine a confidence score§for a hypothesis, certain confidence features are§combined. In this work, the performance of§filler-model based confidence features are§investigated. Five types of filler model are defined:§triphone-network, phone-network, phone-class network,§5-state catch-all model and 3-state catch-all model.§First all models are evaluated in terms of their§ability to correctly tag (miss or hit) recognition§hypotheses. Then the performance of reliable§combinations of these models are evaluated to show§how certain reliable combinations of filler models§could significantly improve the accuracy of the§confidence annotation. Moreover to show the practical§side of the work, an implementation of a real§dialogue management system is described.