| 摘要 |
The method involves estimating (22) probability (p) of presence of speech in a noisy acoustic signal. A spectral covariance matrix (Rn) of noise picked up by microphone sensors is estimated (46). Transfer functions (H) of acoustic channels between a speech source and the sensors are estimated. An optimal linear projector giving a single de-noised combined signal is calculated (48) from the matrix and the functions. The noise is selectively reduced (50) by applying variable gain specific to each frequency band and time frame of the noisy signal based on the probability and the combined signal. |
