EURASIP Journal on Advances in Signal Processing

Table 8 WER [%] on the REVERB challenge dev set, with system combination using both MFCC and PLP features

From: Effectiveness of dereverberation, feature transformation, discriminative training methods, and system combination approach for various reverberant environments

						SIMDATA							REALDATA
		Number of systems				Room 1		Room 2		Room 3		Avg	Room 1		Avg
	ID	GMM	SAT-GMM	SGMM	DNN	Near	Far	Near	Far	Near	Far		Near	Far
1 ch	1)		2			6.00	8.19	7.52	14.37	8.78	18.35	10.54	27.70	30.35	29.03
	2)	2	2			5.31	6.37	6.58	12.62	7.42	16.00	9.05	27.26	29.60	28.43
	3)	4	4			5.33	6.39	6.63	12.67	7.49	15.60	9.02	27.01	29.67	28.34
	4)	4	4	4		5.01	6.34	6.33	12.45	6.87	15.43	8.74	26.64	29.80	28.22
	5)	4	4	4	4	4.67	5.88	6.31	11.93	6.63	14.89	8.39	26.58	28.91	27.75
	6)	2	2	2	2	4.52	5.68	6.29	12.00	6.50	15.06	8.34	26.45	29.80	28.13
8 ch	1)		2			4.72	5.83	5.96	8.92	5.37	8.75	6.59	23.27	28.30	25.79
	2)	2	2			4.72	6.02	5.72	8.26	5.14	8.56	6.40	22.27	26.59	24.43
	3)	4	4			4.72	5.83	5.77	8.21	5.19	8.38	6.35	22.52	26.52	24.52
	4)	4	4	4		4.08	5.16	5.62	7.79	4.80	8.38	5.97	22.40	27.00	24.70
	5)	4	4	4	4	4.18	5.11	5.50	7.74	4.85	8.23	5.94	21.90	26.52	24.21
	6)	3	1	4	2	4.18	5.51	5.50	7.74	4.97	8.43	6.06	21.58	26.32	23.95

For GMM systems, f-bMMI is used, while for SGMM and DNN systems, bMMI is used. The number 2 stands for MFCC and PLP systems, and the number 4 stands for MFCC and PLP systems along with their complementary systems. ROVER 6) uses black-box optimization at the stage of system selection and parameter optimization for ROVER. Italicized data were the best systems in each condition

Back to article page