In this study, we propose to compensate additive noise in the log-spectral domain based on its original non-linear distortion function. We assume the clean speech follows a Gaussian mixture model in the log-spectral domain and noise signal is a single Gaussian distribution. Given any noisy speech observation, we estimate the clean speech by using the original nonlinear distortion function among noise, clean and noisy speech based on the MMSE (minimum mean square error) criterion. The MMSE estimation of clean speech ends up with a complex integral. In this work, we propose an efficient algorithm to use some numerical methods to solve the integral. At last, the estimated clean speech will be mapped from the log-spectral domain into the MFCC domain, and sent to a speech recognizer for the recognition results.

Download paper in PDF format.

The documents distributed by this server have been provided by the contributing authors as a means to ensure timely dissemination of scholarly and technical work on a noncommercial basis. Copyright and all rights therein are maintained by the authors or by other copyright holders, notwithstanding that they have offered their works here electronically. It is understood that all persons copying this information will adhere to the terms and constraints invoked by each author's copyright. These works may not be reposted without the explicit permission of the copyright holder.