Confidence calibration measures to improve speech recognition

Despite a large amount of research efforts in the past, we still believe that robust speech recognition and confidence measure will remain as two most active and influential research topics in speech community for having a researchable future. Researchers have proposed to compute a score (preferably between 0 and 1), called confidence measure (CM), to indicate reliability of any recognition decision made by ASR systems. There are many applications where confidence measures are used to make the right decision. Here we can decide whether the given word or phrase or sentence hypothesis corresponds to an actual occurrence of that event or not. Confidence measures will be used to specify a point of operation in the receiver-operating characteristics (ROC). So in this paper we will see the features of speech signal and how the algorithm of maximum entropy model, deep belief networks can be used to calibrate the confidence score.