. Author manuscript; available in PMC: 2021 Mar 18.

Published in final edited form as: IEEE/ACM Trans Audio Speech Lang Process. 2019 Aug 12;27(11):1839–1848. doi: 10.1109/taslp.2019.2934319

Table III.

ESTOI, PESQ, and ΔSDR scores for speaker separation in anechoic conditions. Single-stage networks are trained with anechoic data. Boldface highlights the best result in each condition.

		ESTOI (%)			PESQ			ΔSDR (dB)
TIR (dB)		−12	−6	Average	−12	−6	Average	−12	−6	Average
Unprocessed		24.6	36.4	30.5	1.35	1.58	1.46
DFN	Mapping	62.4	73.6	68.0	2.35	2.65	2.50	14.15	10.77	12.46
DFN	Masking	63.7	76.2	69.9	2.40	2.75	2.57	15.83	12.94	14.28
LSTM	Mapping	67.9	77.3	72.6	2.47	2.76	2.61	14.60	11.14	12.87
LSTM	Masking	69.1	79.7	74.4	2.52	2.86	2.69	16.26	13.32	14.79
BLSTM	Mapping	71.6	79.9	75.7	2.61	2.87	2.74	15.24	11.70	13.47
BLSTM	Masking	72.0	81.5	76.7	2.62	2.94	2.78	16.77	13.58	15.17