Study Of MMSE Speech Enhancement Algorithm Based On The Short-term Spectral Estimation

Speech signal is the important medium of information dissemination, but in theprocess of transmission, all kinds of noise signal often interfere with the speech signal,it was difficult to get pure speech signal, in order to improve the quality of the speechsignal communication, speech enhancement processing has great application value.Speech enhancement technique not only related with the signal processing theory, butalso with human auditory perception characteristics, noise characteristics. At the sametime, the noise source is very extensive. It take on different characteristics indifferentapplications, which leads to not find a common speech enhancement algorithms forprocessing all kinds of noise, but the specific algorithm applied to be possible undercertain circumstances get a good noise reduction effect.This paper mainly studies based on MMSE speech enhancement algorithm, firstof all, in the beginning of the speech enhancement meaning made a detailedintroduction, meanwhile, its history and development is simply described. Then, inthe second part, it introduces related knowledge of speech enhancement, includingvoice characteristics, the human ear hearing characteristics and noise characteristicsand so on, also introduces the classification of speech enhancement algorithm andsome common speech enhancement algorithm.Although the speech signal is the non-stationary and time-varying signal, thespeech signal has a short-term stability. The speech signal can be divided into severalframes. Each frame of speech signal regards as stable, and then processed. Short-timespectral estimation algorithm for speech enhancement generally include: Spectralsubtraction, Wiener filtering method, the MMSE algorithm. In the third part thesedifferent kinds of speech enhancement algorithm using computer simulationcomparison. Then output SNR with the same input SNR determines enhancing effectof the noisy speech signal. In the short-time spectral estimation of speechenhancement algorithm, the simulation results show that the MMSE algorithm is theoptimal estimation algorithm.The fourth part mainly studies the MMSE speech enhancement algorithm. In thepast, in order to facilitate the calculation and the realization for the MMSE speechenhancement algorithm, we used to assume that speech signal to obey the Gaussiandistribution, but the study found that Super-Gaussian distribution is more close to thedistribution of speech signal. Super-Gaussian distribution to take certain parameterscan get our common distribution model (such as Laplace model, gamma model and soon), also found that these models are better than the Gaussian model. In computer simulation experiments, the main is looking for in the Super-Gaussian model underthe optimal parameters value. It makes the speech enhancement effect best, and that inthe same input SNR cases output SNR is the highest.