Continuous digit speech recognition using subword unit HMM juncture

Abstract

In this paper, a new recognition system is described for digit strings of various length in Korean under fluent speech conditions using subword HMM juncture. A robust automatic recognition system for connected digit strings under spontaneous speech conditions is crucial for a number of applications such as voice dialing and credit card or personal identification numbers. It is thought that the recognition of connected digits is a relatively easy task, but the output of the recognizer must be a very exact transcription of the input speech. The process of recognition is to construct the database of digit string models by concatenating subword clusters not trained on target application through the typing task, avoiding the recording of training data each time the digit changes. Thus, the acoustic segments divide into N subword clusters, and each subword cluster generates the HMM model. The goal is to give users a flexible number recognition that can add new connected digits without collecting a new corpus or training new models, namely that can be easily modified for different applications with minimum amount of training and modification effort.