Signal subspace approach for speech enhancement in nonstationary noises

In this paper, we propose a speech enhancement algorithm that is based on signal subspace approach (SSA) combined with RL noise estimation for nonstationary noises. A signal/noise Karhunen-Loeve transform (SNK) is used to implement signal subspace decomposition for noisy speech signal. RL noise estimation is used for adaptive estimates over time for nonstationary noises. Since the noise covariance matrix for SSA can be recursively updated by RL noise estimation, it results that the proposed algorithm can efficiently suppress the noise while reducing speech distortion. Objective experimental measurements indicated that the proposed algorithm was preferred over other SSA methods when tested with sentences corrupted by nonstationary noises.