The Research On Related Spectrum Attenuation Based On Improved Signal-to-Noise Ratio Estimation

Speech recognition as a new technology for human-computer interaction has made a periodical breakthrough, people have more demands for machine in production and life. This makes that speech recognition has a wide application prospect in the field of medical industry, military industry and information services. Speech recognition offers convenience to people's daily life, but in practical application, the speech signal is polluted by all kinds of noise, which resulting in poor recognition performance. The spectrum attenuation method is a simple and efficient speech denoising algorithm, which has low computation complexity, good denoising effect, and can suppress the background noise effectively. However, the problems existed in spectral attenuation algorithm, including the inaccurate estimation of the priori Signal-to-Noise Ratio(SNR), signal distortion and residual noise, still need to be done further studied. This paper, based on the National Science and Technology Support Project "The integration and application of holographic virtual image technology cooperated real scene system in the culture of Tang dynasty ", carries out research on the part of speech denoising.For the problem that maximum likelihood estimation of SNR algorithm can not effectively eliminate the musical noise, a weighted signal-to-noise ratio estimation algorithm based on multi-band spectral subtraction is proposed. Multi-band spectral subtraction is used to remove the noise, according to the definition of the a priori SNR calculating the priori signal-to-noise ratio, then weighting it with maximum likelihood signal-to-noise ratio, making the final SNR estimation is a combination of both. Simulation results show that the modified approach can improve adaptability to complex environment, eliminate background noise and reduce the speech distortion efficiently.For the problem that Decision-Directed(DD) estimator leads to one-frame delay and in the calculation of posterior SNR using power spectrum leading to musical noise, an improved iterative DD approach to solve one-frame delay problem is proposed. The modified approach uses the amplitude spectrum calculating posterior SNR to reduce the musical noise, and matches each gain function with the noisy speech spectrum at current frame to solve the frame delay problem. The last SNR is obtained according to the definition of the priori SNR iterative. The priori SNR for iterative DD method is closer to the actual SNR, the estimation of SNR is more accurate.For the problem that the inaccuracy of noise estimation in DD estimator leads to residual noise, a noise estimation algorithm using sigmoid function as the coefficient is proposed. By updating the spectrum information of each frame to calculate the posterior SNR, then calculates the sigmoid gain function using the posterior SNR as a parameter. Sigmoid gain function is used to smooth the noise power spectrum that can obtain more smoothing a priori SNR. The experimental results show that the improved algorithm can better track the change of noise, and improve SNR estimation ability of DD estimator.