EURASIP Journal on Advances in Signal Processing

Table 5 Average WER [%] on the REVERB challenge dev set using PLP features

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

			1 ch		8 ch
	Feature		SIMDATA	REALDATA	SIMDATA	REALDATA
Kaldi baseline	PLP	ML	22.96	48.90
derev.			19.84	44.15
CSP+BF+derev.					13.98	42.21
+NLMS					14.97	41.15
GMM	+LDA+MLLT	ML	15.63	40.36	12.13	35.11
	+basis fMLLR		13.70	34.21	10.73	29.21
		bMMI	12.78	33.43	8.94	26.84
		f-bMMI	11.91	30.67	8.10	25.72
		f-bMMI _c	12.20	31.67	8.26	26.30
	+SAT	ML	13.55	36.25	10.17	30.85
		bMMI	11.05	35.63	8.06	28.45
		f-bMMI	10.14	33.29	7.32	26.78
		f-bMMI _c	12.20	31.67	7.61	27.59
SGMM		ML	11.90	32.95	8.43	26.99
		bMMI	10.25	33.10	7.13	26.67
		bMMI _c	10.30	33.14	7.19	27.21
DNN		CE	11.30	31.87	8.75	27.33
		bMMI	9.44	30.19	7.25	26.06
		bMMI _c	9.40	30.13	6.74	26.37

Subscript letter “c” represents the proposed “complementary” system. Italicized data were the best systems in each condition

Back to article page