SAD論文整理

Category	Year	Authors	Title	Features	Classifier/Decision rule	Dataset	noise condition	Comparisons	Performance
Statistical model	IEEE SPL 1999	Jongseo Sohn	A Statistical Model-Based Voice Activity Detection	DFT coefficients, Gsussian modeling	Log-likelihood ratio test (LRT), HMM based hang-over scheme	-	NOISEX-92: vehicle, white, babble; 5dB, 15dB, 25dB	G.729	ROC curves
	IEEE SPL 2005	Javier Ramírez	Statistical Voice Activity Detection Using a Multiple Observation Likelihood Ratio Test	DFT coefficients, Gsussian modeling	Multiple observation LRT	AURORA-3 Spanish SpeechDat-Car (SDC)	distant and close-talking in car environments; 5dB	Sohn’s VAD, G.729, AMR1/2, AFE	ROC curves
	IEEE TASLP2011	Dongwen Ying	Voice Activity Detection Based on an Unsupervised Learning Framework	log=mel energies	GMM	TIMIT	NOISEX-92	Sohn, G.729, AMR	ROC curves
	2015	Tomi Kinnunen	HAPPY Team Entry to NIST OpenSAD Challenge: A Fusion of Short-TermUnsupervised and Segment i-Vector Based Speech Activity Detectors	-	Fusion of 6 VADs	NIST 2015 OpenSAD	-	Sohn, G.729, GMM, rSAD, ivectors	DCF
Deep learning	Interspeech 2016	Ruben Zazo, Google	Feature Learning with Raw-Waveform CLDNNs for Voice Activity Detection	Raw waveform	CLDNN	sythetic, 3800h, balanced	daily life noises, 5~30dB	40dim. log mel energies+DNN, LSTM, CLDNN	ROC curves, FAR, MAR
	IEEE TASLP2016	Xiao-Lei Zhang	Boosting Contextual Information for Deep Neural Network Based Voice Activity Detection	MRCG	Multi-resolution stacking (MRS)+bDNN	AURORA-2(8K, noisy), AURORA-4(16K, clean)	NOISEX-92; -5~20dB	SVM, Zhang13, DNN, bDNN	AUC
	Interspeech 2014	Xiao-Lei Zhang	Boosted Deep Neural Networks and Multi-resolution Cochleagram Features for Voice Activity Detection	MRCG	bDNN	AURORA-4	NOISEX-92; -5~5dB	Sohn, Ramirez05, Ying, SVM, Zhang13	AUC
	IEEE TASLP2013	Xiao-Lei Zhang	Deep Belief Networks Based Voice Activity Detection	Pitch, DFT, MFCC, LPC, PLP, AMS	Deep belif network (DBN)	AURORA-2	-5~10dB	G.729., ETSI Wiener filtering, Sohn, Ramirez05/07, Yu, Shin, Ying, SVM	AUC
	IEEE SPL 2018	Juntae Kim	Voice Activity Detection Using an Adaptive Context Attention Model	MRCG	Attention model+LSTM	TIMIT, self-recorded dataset, HAVIC	NOISEX-92; -5~10dB	HFCL, MSFI, DNN, bDNN, LSTM	AUC
	IEEE ISSPIT 2019	Guan-Bo Wang	A Fusion Model for Robust Voice Activity Detection	Fbank	Fusion of BUT, CRNN, RNN, SOV	OpenSAT19(16K), 160h	various background noise	BUT, CRNN, RNN, SOV	DCF
	APSIPA 2019	Guan-Bo Wang	An RNN and CRNN Based Approach to Robust Voice Activity Detection	Fbank	Fusion of RNN, CRNN	OpenSAT19, OpenSAT17	BUT, RNN, CRNN	DCF
	Interspeech 2019	Ruixi Lin	Optimizing Voice Activity Detection for Noisy Conditions	MFCC, Fbank	DAE, CNN	AISHELL(16K, maually labeled), AURORA-2	self-collected noises; 0~20dB	G.729, SVM, DBN, DDNN	Accuracy
	Interspeech 2015	Qing Wang	A Universal VAD Based on Jointly Trained Deep Neural Networks	MRCG	Jointly learning DNN with speech enhancement	AURORA-4	115 noises types including NOISEX-92; -5~20dB	DNN	AUC
	IEEE ICASSP 2016	Sibo Tong	A COMPARATIVE STUDY OF ROBUSTNESS OF DEEP LEARNING APPROACHES FOR VAD	log-mel energies	Noise-aware training, DNN, LSTM, CNN	AURORA-4, WSJ0	6 noises; 5~20dB	DNN, LSTM, CNN	AUC, EER
	ICMSCE 2018	Jaeseok Kim	Voice Activity Detection based on Multi-Dilated Convolutional Neural Network	MRCG	CNN with multi-dilated convolution	TIMIT	NOISEX-92, sound effect library; -12~10dB	bDNN, RNN, CNN	AUC
	Interspeech 2013	Neville Ryant	Speech Activity Detection on YouTube Using Deep Neural Networks	MFCC	DNN	HAVIC, 65h, web videos	-	GMM	EER
	IEEE ICASSP 2019	Rajat Hebbar	ROBUST SPEECH ACTIVITY DETECTION IN MOVIE AUDIO: DATA RESOURCES AND EXPERIMENTAL EVALUATION	log-mel energies	CNN-TD	Movies	MUSAN, Audioset	CNN, CLDNN	F1, TPR+FPR
	Interspeech 2016	Yuya Fujita	Robust DNN-based VAD augmented with phone entropy based rejection of background speech	Fbank	Acoustic model based DNN, entropy criterion	self-collected, mobile voice search, 1200h	-	DNN	ER
	IEEE ICASSP 2013	Thad Hughes	RECURRENT NEURAL NETWORKS FOR VOICE ACTIVITY DETECTION	PLP	RNN	-	-	GMM+SM	ROC curves
	IEEE ICASSP 2013	Florian Eyben	REAL-LIFE VOICE ACTIVITY DETECTION WITH LSTM RECURRENT NEURAL NETWORKS AND AN APPLICATION TO HOLLYWOOD MOVIES	RASTA-PLP	LSTM	Buckeye (26h), TIMIT	4 noises; -6~25dB	Sohn, Ram05, ARG	AUC, FNR+FPR