System and method for multi class approach for trust modeling in automatic speech recognition system

摘要

A system and method for a multi class approach for confidence modeling in an automatic speech recognition system is presented.The trust model may be trained offline using supervised learning.The decoding module is utilized in the system to generate audio file features in audio data.This feature is used to generate hypothetical speech segments that are compared with known speech segments using the edit distance.The comparison is labeled from one of a plurality of output classes.The label corresponds to the degree of whether the voice is correctly converted to the text.Trained confidence models may be applied in various systems, including interactive voice response systems, keyword spotter and open dialog systems.Diagram