Signal Subsapce Based Speech Enhancement And Implementation

Modern communication systems often cannot work in an idealenvironment now, instead they are always working with different backgroundnoise. Speech enhancement algorithms can reduce or suppress the backgroundnoise using different techniques, in order to make the whole system work wellunder such environment and produce satisfactory results. With today’s diversecommunication applications, research on speech enhancement technology hasimportant and practical significance.Speech enhancement has two main objectives, one is to improve speechquality, and the other is to enhance speech intelligibility. And the target isrelated to different applications. Ideally we want to improve the speechquality and intelligibility at the same time, but in fact they are conflict to eachother, in other words if the quality of speech is improved by some algorithm itwill always lead degradation to the speech. So designing a method to suppressthe noise while only causes a little signal distortion is a difficult challenge.Signal subspace method uses knowledge about linear algebra and matrixtheory to realize noise suppression. The main idea is to decompose the wholesignal vector space into two orthogonal subspaces, which are speech subspaceand noise subspace. By nulling the noise space, we can get the estimatedspeech vectors. The advantage of this method is that it generally notintroduces musical noise like other methods in frequency domain and the balance between signal distortion and remaining noise can be adjusted by anargument in this algorithm. Research on signal subspace noise reductiontechnique in recent years mainly focus on the theoretical study andperformance analysis, there is little work put on the practical systemimplementation.First this paper conducts a deep research on the underlining models andprinciples of signal subspace methods. Also, common estimators andimportant arguments in this algorithm are introduced in this paper. Then, inorder to use this algorithm in real-life applications, we provide some aspectsto improve this method. One way is to use a stable-window sharing techniqueto compute covariance matrix in signal subspace algorithm, by this way weimprove its computation performance. Another way is to use a posteriori SNRcomputation method to make an important argument self-adjustable accordingto different segments in the signal vectors so that the algorithm will leave lessnoise in noise-dominated frames and cause less signal distortion inspeech-dominated frames. Meanwhile, the article introduces a solution to dealwith colored noise problem. At last, we use Matlab platform to realize boththe original algorithm and the improved algorithm, the result shows thatsignal subspace method works better than spectral subtraction algorithm andwith improved techniques used it will be more suitable for real-time speechrecognition application. So we develop an online speech recognition systembased on signal subspace method and demonstrate its user interface.