A Study On Speech Endpoint Detection In High Noisy Environment

Generally, in many application of speech signal processing, the quality of voice has declined because of background noise.In recent years, with the development of application area voice communication, people take more focus on voice communication in background with high noise, especially about voice communication in steel industry with high noise.Due to this reason, how to successfully and precisely detect the presence of voice, how to extract as much original voice data from the corrupted signal and introduce as little distortion as possible has been an more and more important research area.The main task of the thesis is to judge that whether the signal is speech signal or not through analyzing the noisy signal, and then divided the different segments.In this paper, we not only analyzed some basic theory knowledge, and simply introduced the differences between speech signal and noisy signal, and studied signal spectral entropy,AMDF's efforts on speech signal and energy-zero quotient methods, but also improved them individually and make comparative analysis under different contaminating noises. An algorithm that combined the three methods is put forward, and the results give satisfaction on judging whether the signal is speech segment or silent segment.When studying algorithm, we considered that the noises which exist in high noise industry (e.g. steel industry) contain white-noise, pink-noise, destroyerengine-noise, factory1-noise, babble-noise and volvo-noise, and analyze and deal with the signal in frequency domain and time domain separately.The experiments results show that this method that has been improved has better performance in different SNR circumstances than before, and it is suitable for communication in high noise circumstance.