Mandarin Speech Recognition Using Keyword Spotting in a Real-Time System

Abstract

In this paper, we applied the keyword spotting technique to Mandarin speech recognition system based on the semi-continuous Hidden Markov Models (SCHMMS) for automatic interpretation of phone operator. The technique separately deals with keywords and nonkeywords models in training phase. In the testing process, an HMM-based connect word recognition system is used to find the best sequence of keyword, nonkeyword for matching the actual input speech. Finally, we tasked with 35 keywords (names) in a telephone network system and developed an automatic phone operator system for speaker-independent case. The real-time recognizer was implemented on a PC-486 computer enhanced by only one digital signal processing board on which a TMS-320C30 chip operates as CPU. The best recognition accuracy for 35 keywords (names) is 95.72% in the speaker-independent case.