Blind speech dereverberation based on a Bayesian approach to convolutive non-negative matrix factorization.
The algorithm needs a .wav file corresponding to the reverberant signal (isrev = 1), or a clean signal that it will make reverberant if isrev = 0. It also requires model parameters regarding the STFT, stopping criteria and distribution parameters.
For the provided example file, it is recommended to set: B = 1e6, NU = 1, p = 1, Nh = 20, Window size = 512, Window overlapping = 256, stopping parameter = 1e-3, maximum iterations = 20.
A processed sample is provided in the Data link.
Contact: Francisco Ibarrola