Abstract: An unsupervised speech enhancement algorithm suitable for real-time processing in one channel record is proposed, aiming at resolving the prior-information-reliance and real-time processing difficulty in existing enhancement algorithms based on dictionary learning. With the magnitude of noisy speech, it recasts unsupervised background noise modeling problem into sparse, low-rank and noise decomposition. Subsequently, an adaptive noise dictionary which reflects the dynamic noise background is learned in an online fashion by employing incremental nonnegative subspace learning. Finally, frame-by-frame enhancement is conducted with the learnt dictionary, which makes the real-time processing much more convenient. Extensive experiments demonstrate that the presented algorithm outperforms state-of- the-art method such as multi-band spectral subtraction and method based on low-rank and sparse matrix decomposition, especially in terms of noise reduction.