Joint Dereverberation and Separation With Iterative Source Steering

Abstract

We propose a new algorithm for joint dereverberation and blind source separation (DR-BSS). Our work builds upon the IRLMA-T framework that applies a unified filter combining dereverberation and separation. One drawback of this framework is that it requires several matrix inversions, an operation inherently costly and with potential stability issues. We leverage the recently introduced iterative source steering (ISS) updates to propose two algorithms mitigating this issue. Albeit derived from first principles, the first algorithm turns out to be a natural combination of weighted prediction error (WPE) dereverberation and ISS-based BSS, applied alternatingly. In this case, we manage to reduce the number of matrix inversion to only one per iteration and source. The second algorithm updates the ILRMA-T matrix using only sequential ISS updates requiring no matrix inversion at all. Its implementation is straightforward and memory efficient. Numerical experiments demonstrate that both methods achieve the same final performance as ILRMA-T in terms of several relevant objective metrics. In the important case of two sources, the number of iterations required is also similar.

Target Speech Extraction: Independent Vector Extraction Guided by Supervised Speaker Identification

“…where (12) is the orthogonal constraint (OGC) ensuring mutual orthogonality of w k and a k t . A normalization…”

“…The reduced mixture x k,i+1 t is obtained by the least square subtraction of the extracted signal ŝk,i t = (w k,i ) H x k,i t from x k,i t . Let a k,i t be the mixing vector after i deflation steps computed on the tth block via (12). Due to the orthogonality of w k,i and a k,i t , the subtraction is achieved through…”

“…This task is often solved using speech separation methods; all sources present in the mixture are estimated and subsequently the SOI is identified among them. Separation can be performed in a supervised manner using principles of machine/deep learning [1]- [5] or without any prior information about the sources and the mixing using a blind approach [6]- [12].…”

“…Multichannel NMF (MNMF, [9]) extends this concept for analysis of multi-channel mixtures. Independent low-rank matrix analysis (ILRMA, [10]- [12]) unifies the principles of IVA and NMF. Spectral masking methods [19], [20] use the assumption that only one source is dominant at each time-frequency point.…”

Target Speech Extraction: Independent Vector Extraction Guided by Supervised Speaker Identification

“…This task is often solved using speech separation methods; all sources present in the mixture are estimated and subsequently the SOI is identified among them. Separation can be performed either via datadriven techniques deriving their models from large sets of training signals [1]- [5] or via model-based techniques utilizing general statistical assumptions about the sources and the mixing [6]- [12].…”