DNN-Based Frequency-Domain Permutation Solver for Multichannel Audio Source Separation
Fumiya Hasuike, Daichi Kitamura, Rui Watanabe · 2022 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC) · 2022
This paper focuses on frequency-domain blind source separation (BSS) for audio signals. This technique esti-mates frequency-wise source components from an observed spec-trogram. Full-rank spatial covariance analysis and frequency-domain independent component analysis are algorithms com-monly used for this task. Using these methods, however, results in an alignment problem of frequency-wise permutations of the estimated source components. This is known as the permutation problem, which has been addressed for decades and requires a robust and precise permutation solver post-processing. We introduce a permutation solver that uses a deep neural network and predicts the correct source permutations in each frequency. The experimental results demonstrate the validity of the proposed approach and its robustness against the domain of the dataset.