| Speech enhancement extracts the pure speech signal from the mixture as far as possible to improve the quality and intelligibility. From the research status and development results, Speech enhancement algorithm is difficult to estimate not only the problem of noise power, but also to improve the limited SNR in low SNR and non-stationary noise environment. According to the speech itself characteristic, sparse representation can structure original speech and keep speech features to improve the speech quality. For the reason there are many sparse noises in the structured dictionary of K-SVD sparse representation, so sparse representation performs not well under low SNR conditions.In order to solve the above problems, this dissertation researches and designs a strong adaptability and good effect speech enhancement algorithm by applying K-SVD sparse representation. Through the signal subspace algorithm structure noise dictionary, effectively distinguish between noise and speech. Speech dictionary was trained by K-SVD algorithm. The relationship by K-SVD sparse representation speech dictionary reconstruct speech. This algorithm can solve the problems of the noise drown under low SNR of speech, the signal-to-noise ratio is greatly improved. But due to the randomness and mutability of noise, the noise dictionary constructed by subspace, makes it harder to suppress find the characteristics of non-stationary noise than suppress stationary noise.This dissertation utilizes deep neural network(DNN) to improve subspace algorithm to guarantee the powerful capability of constructed noise dictionary. Based on its nonlinear structure, DNN learns the nonlinear relationship between mixed speech and clean speech to enhance the speech results under non-stationary noise conditions.A large number of experimental simulations show that the proposed approaches by DNN and subspace improved K-SVD sparse representation lead to significant improvements of enhanced speech intelligibility and quality. |
PDF Full Text Request