Evaluations on underdetermined blind source separation in adverse environments using time-frequency masking

Jafari, Ingrid; Haque, Serajul; Togneri, Roberto; Nordholm, Sven

doi:10.1186/1687-6180-2013-162

EURASIP Journal on Advances in Signal Processing

Table 5 Source separation results in a reverberant enclosure (cf. Figure 1 ) with background noise

From: Evaluations on underdetermined blind source separation in adverse environments using time-frequency masking

		White noise					Babble noise					Factory noise
Envir.	Cluster	SIR	SINR	SNR	PESQ	OPS	SIR	SINR	SNR	PESQ	OPS	SIR	SINR	SNR	PESQ	OPS
SNR	scheme	(dB)	(dB)	(dB)		(%)
30 dB	HKM	5.27	5.91	20.29	1.34	22.50	5.10	6.23	20.84	1.36	21.48	5.12	5.89	22.53	1.34	24.45
	GMM	6.87	7.71	22.81	1.83	30.15	6.64	7.57	23.52	1.81	21.83	6.34	6.22	22.86	1.71	20.30
	FCM	11.62	11.49	24.66	2.42	35.88	11.28	12.12	25.32	2.50	35.46	11.56	9.45	25.12	2.36	35.63
25 dB	HKM	5.61	5.97	15.90	1.35	22.90	5.83	5.32	16.76	1.34	21.44	5.23	4.73	15.42	1.33	23.92
	GMM	6.60	6.37	20.33	1.81	23.63	6.27	6.07	20.32	1.80	23.91	6.07	5.94	20.31	1.70	20.68
	FCM	9.01	8.76	23.68	2.40	35.10	10.26	10.04	24.05	2.36	35.40	9.42	9.25	24.08	2.30	35.84
20 dB	HKM	5.46	5.69	14.01	1.35	23.18	5.80	5.25	14.30	1.33	23.30	4.98	4.49	13.75	1.28	23.16
	GMM	6.40	5.96	16.67	1.68	23.64	6.15	5.87	18.34	1.67	22.52	6.26	6.01	16.61	1.66	20.70
	FCM	8.05	7.77	21.05	2.34	32.31	10.08	9.64	21.69	2.35	34.64	9.22	8.90	21.36	2.27	33.76
15 dB	HKM	5.77	5.11	12.85	1.36	23.01	4.16	4.86	13.68	1.32	22.03	5.81	5.88	12.46	1.34	22.54
	GMM	6.11	5.35	12.69	1.61	23.51	5.28	5.82	12.35	1.60	21.18	4.70	4.28	13.75	1.59	23.14
	FCM	7.12	6.43	16.93	2.21	28.55	7.18	6.66	17.85	2.42	33.34	8.52	7.59	17.94	2.21	32.91
10 dB	HKM	4.46	4.90	10.03	1.31	23.80	4.72	4.91	10.94	1.30	22.13	4.67	4.47	10.26	1.30	22.44
	GMM	4.01	4.21	9.64	1.58	23.50	5.32	4.61	11.55	1.58	22.16	5.12	4.40	8.79	1.60	22.34
	FCM	7.97	6.39	12.29	1.90	26.59	8.65	7.05	12.39	2.10	27.31	6.97	6.85	12.35	1.96	27.11
5 dB	HKM	3.09	3.08	5.61	1.14	22.45	3.10	3.71	4.71	1.18	21.51	3.09	2.72	6.45	1.17	22.25
	GMM	3.90	3.16	6.26	1.16	22.94	4.65	3.51	5.74	1.40	22.00	4.08	4.91	5.78	1.31	21.05
	FCM	6.73	6.37	7.29	1.90	24.82	6.52	5.96	7.39	1.89	24.13	5.55	6.97	7.15	1.81	23.88
0 dB	HKM	2.16	-0.60	2.30	1.10	22.01	1.47	-0.34	1.73	0.98	21.03	1.04	-1.70	2.12	0.90	22.24
	GMM	1.81	-0.70	1.60	1.04	21.92	-0.46	-0.27	1.70	1.10	22.19	-0.84	-0.64	2.59	1.00	20.57
	FCM	2.60	-0.35	2.54	1.15	24.00	2.72	-0.25	2.38	1.14	23.57	2.58	-0.62	2.51	1.13	22.86
-5 dB	HKM	0.46	-2.74	-2.71	1.09	21.69	-0.49	-3.29	-0.79	0.92	20.56	-1.19	-2.81	0.80	1.07	21.22
	GMM	-0.58	-4.59	-0.43	1.02	21.73	-1.01	-3.29	-0.36	0.91	20.57	-1.20	-2.34	1.64	0.92	20.49
	FCM	0.88	-2.62	-1.67	1.10	21.93	0.78	-0.42	-0.41	1.12	21.00	1.52	-1.82	2.08	1.10	21.91
-10 dB	HKM	-2.54	-5.58	-3.34	0.85	21.05	-2.48	-5.13	-3.10	0.79	19.83	-2.34	-4.01	-3.10	0.97	20.37
	GMM	-1.59	-6.76	-3.71	0.89	21.04	-1.01	-5.64	-1.91	0.90	20.29	-2.13	-5.32	-3.98	0.90	21.05
	FCM	-1.55	-2.51	-3.14	1.00	21.17	-0.64	-4.16	-1.42	1.01	21.01	-1.57	-3.30	-2.13	0.99	21.09

The room reverberation is set to RT₆₀ = 300 ms. The HKM, GMM and FCM clustering algorithms are compared for TF mask estimation using the performance metrics of SIR gain, SINR and SNR as defined in section 4.2. The highest achieved ratio for each acoustic condition is denoted in italics.

Back to article page