EURASIP Journal on Audio, Speech, and Music Processing

Table 6 The comparison of RASTA and the proposed EMD method for every noise condition and every test subset (70 subsets total) of the Aurora 2.0 clean-train tasks

From: Noise-robust speech feature processing with empirical mode decomposition

	Clean training--results
EMD	A					B					C (MIRS)
	Sub	Bab	Car	Exh	Avg	Res	Str	Air	Sta	Avg	Sub	Str	Avg
Clean	98.3	99.5	98.5	98.7	98.8	98.3	98.5	98.5	98.7	98.5	98.5	98.7	98.6
20 dB	96.0	97.1	96.7	95.6	96.4	97.1	96.7	96.5	96.1	96.6	96.2	96.8	96.5
15 dB	93.8	94.0	93.7	91.9	93.4	93.9	93.7	94.1	92.9	93.7	92.7	93.1	92.9
10 dB	86.0	87.4	86.5	82.8	85.7	87.0	87.3	86.6	86.0	86.7	84.6	85.1	84.9
5 dB	73.1	68.2	69.7	66.6	69.4	67.9	71.5	70.1	66.5	69.0	66.1	67.2	66.7
0 dB	43.6	33.7	37.3	41.3	40.0	36.0	42.0	38.2	34.2	37.6	34.4	37.1	35.8
-5 dB	16.3	11.6	12.0	17.1	14.3	12.8	15.9	13.5	11.9	13.5	12.3	15.2	13.8
0-20 dB	78.5	76.1	76.8	75.6	76.8	76.4	78.2	77.1	75.1	76.7	74.8	75.9	75.3
RASTA	A					B					C (MIRS)
	Sub	Bab	Car	Exh	Avg	Res	Str	Air	Sta	Avg	Sub	Str	Avg
Clean	98.8	98.9	99.1	99.2	99.0	98.8	98.9	99.1	99.2	99.0	98.7	98.8	98.8
20 dB	95.7	96.9	96.7	95.2	96.1	95.8	96.6	96.9	97.1	96.6	95.4	96.0	95.7
15 dB	90.1	91.6	89.7	88.8	90.1	91.9	90.3	92.9	90.9	91.5	89.7	90.9	90.3
10 dB	71.4	75.6	64.2	70.4	70.4	79.1	70.0	79.6	72.8	75.4	72.2	72.5	72.4
5 dB	39.8	43.2	30.0	34.6	36.9	50.8	41.3	48.6	40.0	45.2	40.7	40.4	40.6
0 dB	20.6	20.2	17.6	16.1	18.6	24.0	20.3	25.0	20.3	22.4	20.5	20.3	20.4
-5 dB	12.8	10.9	10.1	8.4	10.6	12.2	10.6	13.5	10.5	11.7	12.5	10.8	11.7
0-20 dB	63.5	65.5	59.6	61.0	62.4	68.3	63.7	68.6	64.2	66.2	63.7	64.0	63.9

Back to article page