Speech Enhancement Algorithm Research Base On Short—term Spectrum Estimation

Speech is an important carrier of information. The quality of speech will not only affect the auditory effects, but also can affect the output of all kinds of speech signal processing system. In the real environment, speech signal is hard to avoid being contaminated by noise, so voice quality is decreased. Speech enhancement technique is to remove all kinds of interference noise. Its purpose is to resume the original clean speech signal as much as purity. This paper investigates the two main methods of short-term spectral estimation algorithm, which is the spectral subtraction and minimum mean square error estimation algorithm (MMSE), and improves the two algorithms. Finally, combining these two studies, it proposes an innovation algorithm of geometric spectrum. The main works is as follows:(1) Using the traditional spectral subtraction to do speech enhancement, the result will involve "musical noise". To conquer this bug, we propose a multi-band spectral subtraction. The noisy signal and the estimated noise signal are divided into several frequency bands overlap each other in different frames and different frequency bands. The last, we use each band of the noisy signal and the noise signal to calculate the noise ratio, and use adaptive algorithm to obtain the reduction factor of each band. The new algorithm enhances speech efficiency and the "musical noise" is significantly weakened, and the speech intelligibility is improved.(2) Study the minimum mean square error (MMSE) algorithm based on the maximum likelihood estimation under the log spectral, and purposes a logarithmic spectrum of Estimation Algorithm for MMSE. The approach is to reframe each signal by maximum likelihood estimation, and estimation can be dynamically adjusted by adding the parameter. Relative to the MMSE algorithm, the new algorithm has more noise reduction capabilities, especially for white noise. Residual "musical noise" has declined.(3) On the ideal of the traditional spectral subtraction hypothesis (voice signal and noise are independent), ignore the noisy signal short-time power spectrum of cross-spectral values, as well as bring impossibility to eradicate the "musical noise." To overcome these problems, use geometric to model noisy signal, and obtain a new algorithm which is similar to the minimum mean square error (MMSE) estimate a priori SNR Speech Enhancement Algorithm-geometric spectral subtraction. Experimental results show that the new algorithm not only enhances voice, but also reduces the "musical noise" residues, and voice distortion is minimal.