Deep Recurrent Neural Network-Based Autoencoders for Acoustic Novelty Detection

Table 2

Comparison over the three databases of methods by percentage of -measure (). Indicated prediction (D)elay and average weighted by the # of instances per database. Reported approaches are GMM, HMM, OCSVM, compression autoencoder with MLP (MLP-CAE), BLSTM (BLSTM-CAE), and LSTM (LSTM-CAE), denoising autoencoder with MLP (MLP-DAE), BLSTM (BLSTM-DAE), and LSTM (LSTM-DAE), and related versions of nonlinear predictive autoencoders NP-MLP-CAE/AE/DAE and NP-(B)LSTM-CAE/AE/DAE.

Method	A3Novelty		PASCAL CHiME		PROMETHEUS								Weighted average
					ATM		Corridor		Outdoor		Smart-room
				(%)		(%)		(%)		(%)		(%)	(%)
OCSVM	—	91.8	—	63.4	—	60.2	—	65.3	—	57.3	—	57.4	73.3
GMM	—	89.4	—	89.4	—	50.2	—	49.4	—	56.4	—	59.1	78.7
HMM	—	88.2	—	91.4	—	52.0	—	49.6	—	56.0	—	59.1	78.9
MLP-CAE	0	97.6	0	85.2	0	76.1	0	76.2	0	64.8	0	61.2	84.8
LSTM-CAE	0	97.7	0	89.1	0	78.5	0	77.8	0	62.6	0	64.0	86.2
BLSTM-CAE	0	98.7	0	91.3	0	78.4	0	78.4	0	62.1	0	63.7	87.3
MLP-AE	0	97.2	0	85.0	0	76.1	0	77.0	0	65.1	0	61.8	84.8
LSTM-AE	0	97.7	0	89.1	0	78.7	0	77.9	0	61.6	0	61.4	86.0
BLSTM-AE	0	97.8	0	89.4	0	78.5	0	77.6	0	63.1	0	63.4	86.4
MLP-DAE	0	97.3	0	87.3	0	77.5	0	78.5	0	65.8	0	64.6	86.0
LSTM-DAE	0	97.9	0	92.4	0	79.5	0	78.7	0	68.0	0	65.0	88.1
BLSTM-DAE	0	98.4	0	93.4	0	78.7	0	79.8	0	68.5	0	65.1	88.7
NP-MLP-CAE	4	98.5	5	88.3	1	78.8	2	75.0	1	65.2	5	64.0	86.4
NP-LSTM-CAE	5	98.8	1	92.5	1	78.7	2	74.4	2	64.7	3	63.8	87.7
NP-BLSTM-CAE	4	99.2	3	92.8	2	78.3	1	75.2	2	65.7	2	63.2	88.1
NP-MLP-AE	4	98.5	5	85.9	1	79.0	2	74.6	1	64.7	2	62.8	85.5
NP-LSTM-AE	5	98.7	1	92.1	1	78.1	1	75.0	1	65.0	3	64.4	87.6
NP-BLSTM-AE	4	99.2	2	94.1	3	78.4	2	75.6	1	65.6	1	63.6	88.5
NP-MLP-DAE	5	98.9	5	88.8	1	81.6	1	77.5	1	67.0	4	64.3	87.3
NP-LSTM-DAE	5	99.1	1	94.2	4	80.4	1	76.4	2	66.2	1	65.2	88.8
NP-BLSTM-DAE	5	99.4	3	94.4	2	80.7	3	78.5	2	66.7	1	65.6	89.3