EURASIP Journal on Audio, Speech, and Music Processing

Table 1 AUC (%) comparison between the proposed method and DNN-based VAD methods using speech period candidates and log power spectra as the baseline

From: Enhancement of speech dynamics for voice activity detection using DNN

		AUC (%)—mean ± standard deviation
Noise	SNR (dB)	Proposed	Log power spectra	Speech period candidates
Clean		99.06 ±0.13	98.72 ±0.20	98.10 ±0.39
White	10	97.91 ±0.28	97.51 ±0.49	97.06 ±0.54
	5	97.44 ±0.43	97.27 ±0.48	96.64 ±0.46
	0	96.59 ±0.50	96.14 ±0.76	95.44 ±0.57
	− 5	94.69 ±0.66	93.88 ±1.10	93.40 ±0.60
Babble	10	96.84 ±0.60	96.50 ±0.55	96.19 ±0.68
	5	95.19 ±0.71	94.26 ±0.66	94.59 ±0.92
	0	91.30 ±0.74	88.88 ±0.74	90.42 ±0.53
	− 5	83.20 ±0.87	78.10 ±1.10	81.85 ±0.85
Factory	10	97.25 ±0.39	96.80 ±0.60	96.60 ±0.56
	5	95.96 ±0.43	95.14 ±0.72	95.48 ±0.77
	0	93.18 ±0.46	91.17 ±0.45	92.53 ±0.67
	− 5	85.91 ±0.29	80.49 ±1.54	84.57 ±0.83
Car	10	99.02 ±0.11	98.83 ±0.15	97.60 ±0.45
	5	98.94 ±0.11	98.75 ±0.16	97.37 ±0.45
	0	98.79 ±0.09	98.56 ±0.16	97.02 ±0.41
	− 5	98.40 ±0.05	98.06 ±0.02	96.36 ±0.32
Pink	10	97.79 ±0.39	97.20 ±0.66	96.86 ±0.73
	5	96.82 ±0.59	96.28 ±0.79	95.98 ±0.74
	0	95.26 ±0.70	94.06 ±0.95	94.26 ±0.89
	− 5	91.56 ±1.03	88.01 ±1.54	89.91 ±1.20

The numbers in italics indicate the best results

Back to article page