Weighted Recursive Least Square Filter and Neural Network Based Residual ECHO Suppression for the AEC-Challenge

Abstract

This paper presents a real-time Acoustic Echo Cancellation (AEC) algorithm submitted to the AEC-Challenge. The algorithm consists of three modules: Generalized Cross-Correlation with PHAse Transform (GCC-PHAT) based time delay compensation, weighted Recursive Least Square (wRLS) based linear adaptive filtering and neural network based residual echo suppression. The wRLS filter is derived from a novel semi-blind source separation perspective. The neural network model predicts a Phase-Sensitive Mask (PSM) based on the aligned reference and the linear filter output. The algorithm achieved a mean subjective score of 4.00 and ranked 2nd in the AEC-Challenge.

Joint Acoustic Echo Cancellation and Blind Source Extraction based on Independent Vector Extraction

“…By incorporating the score function normalization into (17) and (20) we obtain the proposed AEC filter update…”

“…This problem has been addressed by Kalman filter-based inference of the model parameters [4,9,10] and machine-learning supported variants [11][12][13]. Besides Kalman filter-based approaches, also blind algorithms originating from independent component analysis (ICA) (see, e.g., [14]) have shown promising results for interference-robust adaptation control [15][16][17][18]. They are particularly interesting as they require no prior training and thus are very robust w.r.t.…”

“…echo cancellers using various statistical source models and update rules [15][16][17][18][19]. Recently, also independent vector analysis (IVA)-based joint estimation of AEC filters and beamforming filters has been investigated [20].…”

Personalized Acoustic Echo Cancellation for Full-duplex Communications

“…Peng et al [7] introduce a time delay estimation (TDE) block besides the adaptive filter to alleviate the task difficulty and propose a gated complex convolutional recurrent neural network (GCCRN) as a post-filter. Zhang et al [8] apply a doubletalk friendly weighted recursive least square (wRLS) filter [9] and use a multi-task gated convolutional frequency-time-LSTM neural network (GFTNN) that predicts the near-end speech activity at the same time. Earlier work on DNN-based acoustic echo cancellation also include [10,11].…”

F-T-LSTM based Complex Network for Joint Acoustic Echo Cancellation and Speech Enhancement

“…Fazel et al [10] designed a deep contextual-attention module with frequency domain NLMS to adaptively estimate features of the near-end speech. Wang et al [11] and Valin et al [12] have also achieved competitive results in the recent AEC-challenge [13]. Zhang and Wang [14] formulated AEC as a supervised speech separation problem, where a bidirectional long-short term memory (BLSTM) network was adopted to predict a mask for magnitude of microphone signal.…”