Speech Recognition Based On Spectral Subtraction In Noisy Environment

Voice, which has long been the most important and basic way of communication in human life, is the most basic way to obtain the necessary information from the living environment. Voice recognition technology is now the ultimate goal of the study is to allow computers to "understand " the human language. Human beings are no longer required to input commands to the computer, and the information interaction between the human and the computer directly depends on the voice commands. The speech recognition system, has been able to achieve high accuracy for clean speech identification, but speech recognition system is now facing a major challenge is the recognition performance in noisy environments. This is due to the noise, has trained speech model and identify noisy speech features between the mismatch, this mismatch will decrease the recognition rate, even can not be identified.Speech recognition technology in noisy environment, the main purpose is to minimize the mismatch between the training model and test speech caused by noise. In order to achieve this goal,to further improve the performance of speech recognition system in noisy environment,it is necessary to pre process and noise reduction processing,as far as possible to filter out noise, improve the signal to noise ratio. This is the most fundamental way to solve the recognition rate of speech recognition system in noise environment.In this paper, the algorithm of speech enhancement is studied, and put forward the improved spectral subtraction algorithm. In the traditional spectral subtraction algorithm,based on the silent smoothing, noise power spectrum can be obtained. The traditional spectral subtraction is simple, good real-time performance, but eventually the noise reduction effect and can not achieve the requirements of noise suppression. Therefore,in order to futher improve the spectral subtraction algorithm for anti noise performance in noise environment,this paper conducted a series of improvements on spectral subtraction.