Low-delay noise estimation based on spectrum ripples and minimum statistics in adverse environments

This paper proposes a new noise estimation algorithm to reduce the estimation delays under highly non-stationary noise conditions. Since the harmonic ripples appeared in the spectrogram are valuable for human to localize the speech presence, based on the characteristics of these ripples, we propose a novel energy independent feature to detect the changing noise. If noise is present, the noise floors of the traditional minimum statistics (MS) are forced to update to follow the noise change. This scheme can also prevent the false rise of noise floors of MS during long speech presence. The performance of the proposed algorithm is evaluated by qualitative results and overall objective measures. Better performances are achieved compared with other noise estimation algorithms.