Performance Evaluation of Independent Low-rank Matrix Analysis for Short Signals

Taishi Nakashima , Robin Scheibler (1) , Yukoh Wakabayashi , Nobutaka Ono (2)

1 TMU - Tokyo Metropolitan University [Tokyo]
2 UTokyo - The University of Tokyo

In this paper, we evaluate the performance of independent low-rank matrix analysis (ILRMA) for short signals. ILRMA is a state-of-the-art blind source separation (BSS) technique based on the assumptions that sources are statistically independent, and their spectrograms are approximately expressed as low-rank matrices. Because ILRMA estimates many parameters such as demixing matrices, spectral bases, and source activations, it needs a sufficient-length observation for the stable estimation. Then, the performance of ILRMA could degrade when the available signals are short. For improving this, we apply the latest ILRMA algorithms to a short mixture and investigate the dependence of the performance on the signal length.

Mots clés: blind source separation; independent vector analysis; non-negative matrix factorization; independent low-rank matrix analysis

Domaines: Acoustique [physics.class-ph]; Vibrations [physics.class-ph]

Taishi Nakashima, Robin Scheibler, Yukoh Wakabayashi, Nobutaka Ono. Performance Evaluation of Independent Low-rank Matrix Analysis for Short Signals. Forum Acusticum, Dec 2020, Lyon, France. pp.837-840, ⟨10.48465/fa.2020.0720⟩. ⟨hal-03235358⟩