Single-channel speech enhancement based on non-negative matrix factorization and online noise adaptation

In this paper, we demonstrate a simulator for real-time speech enhancement based on a non-negative matrix factorization (NMF) technique. In particular, we propose an online noise adaptation method in an NMF framework, which is activated during non-speech intervals and used for adapting noise bases for NMF. Thus, incoming noisy speech is decomposed by using such adapted noise bases and universal speech bases that can be developed through training with examples of speech data. It is shown from the experiments that the proposed method improves speech separation performance and perceptual speech quality.