A family of 3GPP-standard noise suppressors for the AMR codec and the evaluation results

The paper presents a family of 3GPP-standard noise suppressors and the evaluation results. The family consists of a high-quality version and a low-complexity version. These noise suppressors are based on the MMSE STSA (minimum mean square error short time spectral amplitude) algorithm originally proposed by Y. Ephraim and D. Malah (see IEEE Trans. Acoust., Speech, Sig. Processing, vol.ASSP-32, no.6, p.1109-21, 1984). To meet the 3GPP requirements with better speech quality, weighted noise estimation, synthesis windowing, and pseudo noise injection are incorporated. Weighted noise estimation enables continuous noise estimation even in the speech period by using weighted noisy speech. The weight is controlled such that a higher estimated SNR gives a smaller weight. A synthesis window function is applied between inverse transform and overlap-add processing for smooth transition at frame boundaries. Pseudo noise injection, which is not available in the low-complexity version, modifies the spectral gain based on its nonlinearity. The whole family satisfies all the 3GPP requirements. Results of a full set of evaluations specified by 3GPP are presented for the high-quality version.