Fig. 5:
Comparison of the proposed two-stage network with target-dependent and talker-independent speaker separation. All networks use BLSTM and test TIR is −12 dB. (a) ESTOI scores in simulated RIR conditions, (b) ESTOI scores in recorded RIR conditions, (c) PESQ scores in simulated RIR conditions, (d) PESQ scores in recorded RIR conditions, (e) ΔSDR scores in simulated RIR conditions, and (f) ΔSDR scores in recorded RIR conditions. Error bars depict the standard deviation.
