Detection of speech signal in strong ship-radiated noise based on spectrum entropy

Comparing the frequency spectrum distributions calculated from several successive frames, the change of the frequency spectrum of speech frames between successive frames is larger than that of the ship-radiated noise. The aim of this work is to propose a novel speech detection algorithm in strong ship-radiated noise. As inaccurate sentence boundaries are a major cause in automatic speech recognition in strong noise background. Hence, based on that characteristic, a new feature repeating pattern of frequency spectrum trend (RPFST) was calculated based on spectrum entropy. Firstly, the speech is detected roughly with the precision of 1 s by calculating the feature RPFST. Then, the detection precision is up to 20 ms, the length of frames, by method of frame shifting. Finally, benchmarked on a large measured data set, the detection accuracy (92 %) is achieved. The experimental results show the feasibility of the algorithm to all kinds of speech and ship-radiated noise.