Intra-frame cepstral sub-band weighting and histogram equalization for noise-robust speech recognition

EURASIP Journal on Audio, Speech, and Music Processing

Table 9 The recognition accuracy results (%) of various forms of WS-HEQ for different test sets

Method		Set A	Set B	Set C	Average
${WS-HEQ}_{I}^{(1)}$	α=0.6	83.36	85.37	83.89	84.27
	Optimal α	83.86	85.56	84.52	84.67
${WS-HEQ}_{I}^{(2)}$	α=0.6	82.29	83.22	82.82	82.76
	Optimal α	83.04	84.08	83.32	83.51
${WS-HEQ}_{I}^{(3)}$	α=0.5	83.57	85.15	83.93	84.27
	Optimal α	83.86	85.46	84.43	84.62
${WS-HEQ}_{I}^{(4)}$	α=0.7	82.88	84.70	82.78	83.59
	Optimal α	83.52	84.86	83.85	84.12
${WS-HEQ}_{II}^{(1)}$	α=0.6	84.13	86.16	84.39	84.99
	Optimal α	84.47	86.39	84.57	85.26
${WS-HEQ}_{II}^{(2)}$	α=0.6	83.54	85.75	83.83	84.48
	Optimal α	84.25	86.04	84.52	85.02
${WS-HEQ}_{II}^{(3)}$	α=0.7	83.25	85.10	83.50	84.04
	Optimal α	83.84	85.85	84.05	84.69
${WS-HEQ}_{II}^{(4)}$	α=0.6	82.30	83.45	82.90	82.88
	Optimal α	82.92	84.18	83.17	83.47

These results are obtained by using (1) the scaling factor α listed in Table 1 (2) the scaling factor α that achieves the optimal recognition accuracy with respect to the individual noise type and level (SNR), both of which are for different Test Sets while averaged over 5 SNR conditions (20 dB to 0 dB) as to the Aurora-2 database.