An algorithm for single channel enhancement of noisy speech has been proposed. The algorithm operates without any knowledge or assumption of noise parameters and reduces the noise in the temporal domain rather than the spectral domain using a non-linear and automatically adjustable gain function for multi-band dynamic range compression. The gain function is deduced from the temporal envelope of each frequency-band and compresses the frequency regions where speech is absent. Objective and blind subjective tests show that the proposed algorithm performs better than three benchmark algorithms, reduces background noise (car and babble), and introduces very little musical noise.
The test signals used are available in the table below:
|
Car 0 dB (C0Noisy) |
Car -5 dB (C-5Noisy) |
Babble 0 dB (B0Noisy) |
Babble -5 dB (B-5Noisy) |
|
|
Wiener filtering |
||||
|
Spectral subtraction |
||||
|
Modulation Filtering |
||||
|
Proposed algorithm |
