RESUMO
This paper addresses the problem of noise reduction in the time domain where the clean speech sample at every time instant is estimated by filtering a vector of the noisy speech signal. Such a clean speech estimate consists of both the filtered speech and residual noise (filtered noise) as the noisy vector is the sum of the clean speech and noise vectors. Traditionally, the filtered speech is treated as the desired signal after noise reduction. This paper proposes to decompose the clean speech vector into two orthogonal components: one is correlated and the other is uncorrelated with the current clean speech sample. While the correlated component helps estimate the clean speech, it is shown that the uncorrelated component interferes with the estimation, just as the additive noise. Based on this orthogonal decomposition, the paper presents a way to define the error signal and cost functions and addresses the issue of how to design different optimal noise reduction filters by optimizing these cost functions. Specifically, it discusses how to design the maximum SNR filter, the Wiener filter, the minimum variance distortionless response (MVDR) filter, the tradeoff filter, and the linearly constrained minimum variance (LCMV) filter. It demonstrates that the maximum SNR, Wiener, MVDR, and tradeoff filters are identical up to a scaling factor. It also shows from the orthogonal decomposition that many performance measures can be defined, which seem to be more appropriate than the traditional ones for the evaluation of the noise reduction filters.