Endpoint Detection Algorithm For Speech Signal In Low SNR Environment

In the field of speech signal recognition,the detection of speech signal endpoint is critical.The purpose of the endpoint detection is to determine the start and end points of the sampling speech signal,so that the power consumption of the speech signal processing can be reduced and the system performance can be improved.Endpoint detection technology is widely used today,such as in voice conversation tools,wearable devices,artificial intelligence machines and many other applications.In recent years,there have been many practical and efficient endpoint detection algorithms,which can be roughly divided into two categories.One is based on the feature parameter extraction of the frequency domain or time domain of digital signals,and the other is based on pattern matching.At present,the research on speech signal endpoint detection technology under the condition of no-noise interference is very mature,but the detect effect will be significantly reduced in low signal noise ratio environment,so it is of great importance to find an accurate and efficient detection algorithm when the SNR is low.The paper first summarizes the traditional algorithms based on feature parameters in time domain and frequency domain.Through evaluation of the detection performance under different noise intensities,it can be concluded that the traditional detection algorithm has high detection accuracy and computational efficiency when there is no noise.However,for low SNR environments,the detection performance of traditional algorithms is rapidly degraded,and the signal endpoints cannot be accurately found.Therefore,for a low SNR environment,the signal can be first denoised using a speech signal denoising algorithm,and then the denoised signal can be used for endpoint detection.Based on this idea,this paper proposes an enhanced joint spectral subtraction and variance method and a joint multi-window spectrum reduction and energy entropy ratio method.These two algorithms effectively combine the denoising and detection processing of speech signals to improve the detection performance.The detection principle of the joint enhanced spectral subtraction and variance method proposed in this paper is to first use enhanced spectral subtraction to denoise,and this method introduces a spectral subtraction factor,so that the spectral subtraction can be better adapted to noise.The variance of the signal denoised by the enhanced spectral subtraction method has obvious characteristics at the boundary point of the speech signal.Therefore,the basic variance method is used for endpoint detection.The joint multi-window spectrum reduction and energy entropy ratio method is to use multi-window spectrum estimation reduction method to denoise the noisy speech signal.This method uses multiple mutually orthogonal window functions to obtain the spectral value,which effectively reduces the experimental errors.The entropy ratio method is more resistant to noise than other conventional endpoint detection algorithms.In order to verify the performance and characteristics of the two new algorithms proposed in this paper,we apply these two new algorithms to audio recognition for experimental analysis.The pure speech data used in this experiment is a natural human voice data set,and noise signal data is from the Noisex-92 noise dataset.The noisy speech signal data sets for experiments are obtained by adding several kinds of additive noise signals respectively to the pure speech signal.By setting different signal-to-noise ratios and adding 'different types of noise,their detection results are observed and analyzed by comparison.Based on the experiment results,it can be concluded that the two new algorithms proposed in this paper have good adaptability to different noise and low SNR environment,and can effectively detect the endpoints of the speech signal.