Speech enhancement using a-priori information with classified noise codebooks

This paper focuses on the estimation of short-term linear predictive parameters from noisy speech and their subsequent use in waveform enhancement schemes. We use a-priori information in the form of trained codebooks of speech and noise linear predictive coefficients. The excitation variances of speech and noise are determined through the optimization of a criterion that finds the best fit between the noisy observation and the model represented by the two codebooks. Improved estimation accuracy and reduced computational complexity result from classifying the noise and using small noise codebooks, one for each noise class. For each segment of noisy speech, the classification scheme selects a particular noise codebook. Experimental results show good performance, especially under non-stationary noise conditions. Listening tests confirm that the new method outperforms conventional speech enhancement systems.