Speech Signal Detection Under Condition Of Low SNR

The original speech signal processing uses time domain parameters directly and spectrum parameters are applied to judge to achieve the better performance of the speech quality.But considering the noise characteristics are strong and non-stationary in actual situation,the worse robustness leads system performance to be more sensitive to signal-to-noise ratio.Firstly,this thesis mainly combs the relevant home and abroad literatures,which from the view of speech signal detection application,practical value and research status under low signal-to-noise condition.It also analyzes the complexity of the speech signal in strong noise environment model to summary the key issues and strong non-stationary noise conditions in speech signal detection.Secondly,a voice endpoint detection algorithm is proposed by short-term signalto-noise ratio,which is based on adaptive threshold and adaptive decision.The shortterm zero-crossing rate and adaptive decision is added to judge the final detection result.Then,in terms of the speech signal complex structure under low signal-to-noise ratio condition,a spectral subtraction algorithm based on sub-band spectral entropy with adaptive learning is proposed.The algorithm divides the noisy speech into several sub-bands and adaptively weights them respectively,and utilize them to estimate the noise spectrum energy.Finally,the lab results show that the adaptive threshold endpoint detection algorithm is able to detect the endpoints between the spoken and non-speech segments effectively,which in the environment with stationary noise and non-stationary noise under different signal-to-noise ratios.The algorithm accuracy and robustness are superior to the original endpoint detection algorithm.In addition,the adaptive sub-band spectral entropy speech enhancement algorithm can achieve a better performance and the speech quality improvement is remarkable from different real noise sources.