Robust speech detector for speech recognition applications

The speech detector takes the significant role in the practical speech recognition system. In this paper, we propose a novel finite state machine based speech detector (FSD) featured with real-time merit. The input of the FSM comes from a frame-energy detector on the principle of logarithm maximum likelihood as well as a zero-crossing detector. Experimental results based on both an isolated word recognition system and a continuous speech recognition system prove that the FSD is feasible and robust.