Online independent vector analysis (IVA) based on auxiliary technology is effective to separate audio source in real time. However, the separated signal may contain residual interference noise because the source model of IVA lacks flexibility and cannot treat the specific harmonic structures of sources. This paper presents a real-time IVA method where the amplitude spectrum of separated signal is modeled by semi-supervised nonnegative matrix factorization (SSNMF). Using the pre-trained basis matrix which contains source structures, we can extract the target source from the separated signal in real time. The advantage of the proposed method is that the extracted source can provide a more accurate variance than the separated signal and hence the proposed method can obtain a better separation performance than the oracle IVA. Experimental results in speech denoising task show the effectiveness and the robustness of the proposed method with different types of noise.
Cite as: Wang, T., Yang, F., Zhu, R., Yang, J. (2021) Real-Time Independent Vector Analysis Using Semi-Supervised Nonnegative Matrix Factorization as a Source Model. Proc. Interspeech 2021, 1842-1846, doi: 10.21437/Interspeech.2021-146
@inproceedings{wang21p_interspeech,
author={Taihui Wang and Feiran Yang and Rui Zhu and Jun Yang},
title={{Real-Time Independent Vector Analysis Using Semi-Supervised Nonnegative Matrix Factorization as a Source Model}},
year=2021,
booktitle={Proc. Interspeech 2021},
pages={1842--1846},
doi={10.21437/Interspeech.2021-146}
}