Efficiency evaluation of subspace-based spectral subtraction based on iterative eigenvalue analysis in real environments

In real environments, the recorded speech signal is much affected by unwanted noise. Therefore, it is necessary to reduce the unwanted noise from the recorded noisy signal. The spectral subtraction (SS) and the flooring processing-improved SS (F-SS) have been proposed to achieve that. The F-SS iteratively estimates the clean speech signal by utilizing the SS. However, the F-SS generates the distortion in the noise-reduced signal although it can reduce the unwanted noise. In this paper, we propose the subspace-based spectral subtraction (S-SS) to reduce the distortion from the noise-reduced signal. The proposed S-SS performs the eigenvalue analysis with multiple noise-reduced signals by the SS. The proposed S-SS acquires multiple noise-reduced signals by the SS under the various conditions of noise estimation. The subspace of speech component is calculated from multiple noise-reduced signals by the eigenvalue analysis. The proposed S-SS then acquires the noise-reduced signal which is reduced the distortion by using the subspace of the speech component. The proposed S-SS can simultaneously reduce the unwanted noise and the distortion of the observed signal by iteratively performing these processes. As a result of objective experiments with signal-to-distortion ratio (SDR), we confirmed that the proposed S-SS can reduce the distortion of the noise-reduced signal.