The past tens of years has witnessed great advance in speech technologies, and many of them have already been widely used in our daily life. However, in the real world conditions, the performance of most speech system always degrades significantly due to the presence of environment noise. The motivation of this paper is to give an effective solution to this problem by means of speech enhancement. We focused our speech enhancement research in a case where a noise reference is available and discussed adaptive noise cancellation and its application in speech recognition system. The author's work presented in this paper can be summarized as follows: 1. Analyzed some typical mono-channel speech enhancement techniques that do not use noise reference, pointed out both their advantages and disadvantages, and then suggested that dual-channel is a better solution. 2. Studied several existing adaptive noise cancellation algorithms (including LMS algorithm, RLS algorithm and the subband method). In order to achieve faster adaptation convergence speed, an improved NLMS algorithm based on filter weight contour is proposed, and satisfactory result has been observed in experiments. In order to solve the nonlinear problem, a method of compensating non-linearity in adaptive noise cancellation is also proposed in this paper, and its effectiveness is proved by experiment simulation. 3. Analyzed, designed, and realized a TV voice command control system that utilizes both adaptive noise cancellation and speech recognition. Introduced its hardware and software design, and the main algorithms used. A refined filtered energy based speech endpoint detection algorithm is also introduced. 4. Evaluated our speech enhancement method based on the above system, and compared it with other algorithms.
修改评论