Consistent Wiener filtering: generalized time-frequency masking respecting spectrogram consistency

Jonathan Le Roux (1) , Emmanuel Vincent (2) , Yuu Mizuno (3) , Hirokazu Kameoka (1) , Nobutaka Ono (3) , Shigeki Sagayama (3)

1 NTT Communication Science Laboratories
2 METISS - Speech and sound data modeling and processing
3 UTokyo - The University of Tokyo

Wiener filtering is one of the most widely used methods in audio source separation. It is often applied on time-frequency representations of signals, such as the short-time Fourier transform (STFT), to exploit their short-term stationarity, but so far the design of the Wiener time-frequency mask did not take into account the necessity for the output spectrograms to be consistent, i.e., to correspond to the STFT of a time-domain signal. In this paper, we generalize the concept of Wiener filtering to time-frequency masks which can involve manipulation of the phase as well by formulating the problem as a consistency-constrained Maximum-Likelihood one. We present two methods to solve the problem, one looking for the optimal time-domain signal, the other promoting consistency through a penalty function directly in the time-frequency domain. We show through experimental evaluation that, both in oracle conditions and combined with spectral subtraction, our method outperforms classical Wiener filtering.

Domaines: Traitement du signal et de l'image [eess.SP]; Traitement du signal et de l'image [eess.SP]

Jonathan Le Roux, Emmanuel Vincent, Yuu Mizuno, Hirokazu Kameoka, Nobutaka Ono, et al.. Consistent Wiener filtering: generalized time-frequency masking respecting spectrogram consistency. 9th Int. Conf. on Latent Variable Analysis and Signal Separation (LVA/ICA), Sep 2010, Saint-Malo, France. pp.89--96. ⟨inria-00544072⟩