Đang chuẩn bị liên kết để tải về tài liệu:
Artificial Mind System – Kernel Memory Approach - Tetsuya Hoya Part 10

Tham khảo tài liệu 'artificial mind system – kernel memory approach - tetsuya hoya part 10', kỹ thuật - công nghệ, cơ khí - chế tạo máy phục vụ nhu cầu học tập, nghiên cứu và làm việc hiệu quả | 106 6 Sensation and Perception Modules Therefore it is considered that the respective stereophonic speech components Si k i 1 2 are generated from one speech source using two sufficiently long filters hj and in reality the stereophonic speech components sj k are strongly correlated with each other. Then the objective here is to eliminate both the noise components n1 k and n2 k from the corresponding observation x1 k and x2 k . For the simulation examples given here the length of the analysis matrix was fixed to 32 for an SSP whilst for the DASE the standard normalised least mean square NLMS algorithm see e.g. Haykin 1996 was used to adjust the adaptive filter coefficients. For each adaptive filter the learning constant was chosen as 0.5 and the filter lengths were 51 the latter selection of which neither precedence effect or alternatively the Haas effect nor echo effect was not considered to occur Hugonnet and Walder 1998 . Moreover the scalar constants cj were empirically adjusted to 0.1 which moderately suppressed the distortion and satisfied a good trade-off between a reasonable stereophonic image compensation and signal distortion. Figure 6.5 a shows the left and right channel signals of a real speech of the sentence Pleasant zoos are rarely reached by efficient transportation uttered in English by a male speaker b the noisy speech assuming the input SNR 3 dB c the enhanced speech obtained from nonlinear spectral subtraction NSS algorithm Martin 1994 Xie and Van Compernolle 1996 Gustafsson et al. 1999 Martin 2001 Gustafsson et al. 2003 which is one of the classical and most commonly used methods for single-channel noise reduction applied independently to each channel d the enhanced speech by a single stage SSP with the data-reusing scheme given in 6.8 Hoya et al. 2003b Hoya et al. 2005 and e the enhanced speech by the combined SSP and DASE method respectively. For the simulation example shown in Fig. 6.5 the noise components n1 k and n2 k are assumed to be two