Surrogate Source Model Learning for Determined Source Separation

Abstract

We propose to learn surrogate functions of universal speech priors for determined blind speech separation. Deep speech priors are highly desirable due to their superior modelling power, but are not compatible with state-of-the-art independent vector analysis based on majorization-minimization (AuxIVA), since deriving the required surrogate function is not easy, nor always possible. Instead, we do away with exact majorization and directly approximate the surrogate. Taking advantage of iterative source steering (ISS) updates, we back propagate the permutation invariant separation loss through multiple iterations of AuxIVA. ISS lends itself well to this task due to its lower complexity and lack of matrix inversion. Experiments show large improvements in terms of scale invariant signal-to-distortion (SDR) ratio and word error rate compared to baseline methods. Training is done on two speakers mixtures and we experiment with two losses, SDR and coherence. We find that the learnt approximate surrogate generalizes well on mixtures of three and four speakers without any modification. We also demonstrate generalization to a different variation of the AuxIVA update equations. The SDR loss leads to fastest convergence in iterations, while coherence leads to the lowest word error rate (WER). We obtain as much as 36 % reduction in WER.

音源分離技術の基礎と動向

“…( 1 ) 音源分離における空間フィルタをニューラルネットワー クによって推論 (112)， (113) ( 2 ) ブラインド音源分離の音源モデルへの深層学習の導 入 ( 114) - (120) ( 3 ) 位相差などの複数マイクロホンの信号から得た情報を ニューラルネットワークの特徴量として利用 (103)， (121) ( 4 ) ビームフォーミングで用いる共分散行列をニューラル ネットワークを介して求めた時間周波数マスクを用いて 推定 (122)…”

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

“…Let a k,i t be the mixing vector after i deflation steps computed on the tth block via (8). Due to the orthogonality of w k,i and a k,i t , the subtraction is achieved through…”

“…The recovered frequency components have a random order and all components corresponding to the wide-band source need to be identified in order to reconstruct it in the time-domain. To alleviate this drawback, the independent vector analysis (IVA, [7], [8]) has been proposed. It binds together the frequency components corresponding to a single source using higherorder dependencies among them.…”

Meta-AF: Meta-Learning for Adaptive Filters

“…Recently, it was shown that DNNs can estimate statistics to control step-sizes [61], [62] or estimate entire updates [63] for a single-channel AEC. Similarly, past work has used DNNs to predict updates for the internal statistics of multi-channel beamformers [64] and to learn source-models for multi-channel source separation [65]. These works differ from hybrid approaches in that they leverage neural networks to update or control AFs directly and thus focus on improving the performance of AFs themselves.…”