ITERATIVE SINUSOIDAL-BASED PARTIAL PHASE RECONSTRUCTION IN SINGLE-CHANNEL SOURCE SEPARATION

 

Audio demonstration of results obtained for the accepted paper to INTERSPEECH2013

Mario Kaoru Watanabe and Pejman Mowlaee

Single-channel source separation is imposed as the problem of finding both amplitude and phase spectra of the individual underlying sources given their single-channel recorded mixture as the only observation. While many previous methods suggest different approaches to provide a better estimation for the magnitude spectra of the individual sources, most of these methods directly pass the mixture phase spectrum unaltered to the signal reconstruction stage. Recently in [1-3], we studied the impact of phase in two stages: signal parameter estimation [3] and signal reconstruction [1-2]. For information available here.

    In this work, we propose a sinusoidal-based iterative phase estimation algorithm to estimate the phase of the individual sources, given their Wiener filtered magnitude spectra. Our experimental results reveal that the proposed method compared to other state-of-the-art phase estimation algorithms balances a tradeoff between improved signal-to-distortion and signal-to-artifact performance without much noticeable degradation in signal-to-interference performance, for both oracle and quantized Wiener filtered magnitude spectra.

 

Quantized Wiener filtered magnitude spectra (SQNR = 30dB)

Original mixture
Wiener mixture phase
PPR
Proposed

 

References
[1] P. Mowlaee, R. Saeidi, "On the phase importance in single-channel speech enhancement", in Proceeding of IEEE Int. Conf. Acoustics, Speech, Signal Processing, May. 2013, Vancouver, Canada, In Press.
[2] P. Mowlaee, R. Saeidi, R. Martin, ", In Proceeding of 13th Annual Conference of the International Speech Communication Association,(INTERSPEECH), Portland, USA, Sept., 2012.
[3] P. Mowlaee, R. Martin, "On Phase Importance in Parameter Estimation for Single-channel Source Separation", In Proceeding of International Workshop on Acoustic and Signal Enhancement (IWAENC), Aachen, Germany, Sept., 2012.