Research On Speech Enhancement Algorithm Based On Wavelet Packet Adaptive Threshold

Speech enhancement is a technology that extracts the useful speech and filters noise from speech which is polluted by noise, and it is a research focus of speech signal processing technology; the purposes of speech enhancement are to get rid of background noise and improve the quality of speech and intelligibility, making hearer accept the enhanced speech at the same time can better understand speech, while these purposes usually can’t have your cake and eat it, we should stay focused according to the specific application. This paper first elaborates the basic theory of wavelet transform and the basic principle of speech enhancement algorithm based on the wavelet packet transform, and makes simulation experiment.Aiming at enhanced speech by using the traditional soft and hard threshold functions wavelet packet transform speech enhancement algorithm has speech distortion, this paper proposes new threshold function for wavelet packet speech enhancement algorithm. The new threshold function does improvement when wavelet packet coefficients are greater or less than the given threshold, on the one hand the threshold function flexibly combines the soft and hard threshold functions; on the other hand, it uses a nonlinear function to revise the part of tasking zero value directly in the traditional threshold function, and makes the threshold function achieve the smooth transition. Simulation experimental results show that comparing with traditional method, the new algorithm can get better enhancement effect in white Gaussian noise and colored noise background, and reduce the speech distortion. Noise is changeable, so fixed threshold wavelet packet speech enhancement algorithm causes the speech distortion, aiming at this problem, speech enhancement algorithm based on wavelet packet adaptive threshold is proposed, main content includes the following two aspects:1. This paper proposes a kind of adaptively threshold that is revised by the posteriori signal-to-noise ratio (SNR) of wavelet packet nodes, the noise power is estimated by the improved non-stationary noise estimation algorithm that combines with masking effect, which ensures that posterior SNR of nodes is calculated accurately. The adjacent frame wavelet packet threshold which changes with the scale is smoothed by the sigmoid function with the parameter of posteriori SNR, which ensures the continuity of threshold. Furthermore, this threshold is adaptively revised further by the exponential posterior SNR, which reduces speech distortion.2. This paper improves Bark scale wavelet packet decomposition structure, which can better simulate the human ear auditory perception characteristics. Simulation experimental results show that the proposed algorithm improves the SNR, segment SNR (SSNR) and perceptual evaluation of speech quality (PESQ) of the enhanced speech, and reduces the speech distortion; compared to the fixed threshold wavelet packet speech enhancement algorithm, the proposed algorithm has better enhanced effect.In order to make the adaptive threshold wavelet packet speech enhancement algorithm to achieve the best enhancement effect, subspace speech enhancement algorithm that can give overall consideration between speech denoising and distortion is introduced into the adaptive threshold wavelet packet method, this paper proposes combined with improved subspace of adaptive threshold wavelet packet speech enhancement algorithm. This algorithm uses the adaptive threshold wavelet packet method to do enhance for noisy speech which is done KL transform initially, and uses the Recursive least squares algorithm (RLS) to estimate noise in order to correct the method of estimating noise characteristic value in the traditional subspace method, and overcomes the problem that the traditional subspace speech enhancement algorithm causes speech feature value estimation deviation easily.Simulation experimental results show that, using the advantage of subspace method, this new method could further improve the enhanced effect of adaptive threshold wavelet packet method, especially in low SNR, the increase of the SNR and SSNR of enhanced speech is greater.