List your company
ResearchFormerly WizWand

Speech Recognition on Hub5'00 SWB (test)

6.7WERResNet+LSTM1+LSTM2
5.0416.24527.4538.655Dec 17, 2014Sep 6, 2015May 27, 2016Feb 15, 2017Nov 6, 2017Jul 28, 2018Apr 18, 2019
Updated 8mo ago

Evaluation results

MethodLinks
ResNet+LSTM1+LSTM2Type=Frame-level score fusion2017.036.7
ResNet+LSTM2Type=Frame-level score fusion2017.036.8
LAS + SpecAugment (SM)Architecture=LAS, Policy=SM, Language Model=Yes2019.046.8
LAS + SpecAugment (SS)Architecture=LAS, Policy=SS, Language Model=Yes2019.047.1
LSTM2 (Feat. fusion)Type=Individual model2017.037.2
LAS + SpecAugment (SM)Architecture=LAS, Policy=SM, Language Model=No2019.047.2
LAS + SpecAugment (SS)Architecture=LAS, Policy=SS, Language Model=No2019.047.3
LSTM1 (SA-MTL)Type=Individual model2017.037.6
ResNetType=Individual model2017.037.6
This systemAM training data=SWB+Fisher+CH2015.058
hybridLM=LSTM, label unit=CDp2018.058.3
Zeyer et al. [24]Architecture=HMM, Language Model=Yes2019.048.3
Hadian et al.Architecture=HMM, Language Model=Yes2019.049.3
LF MMILM=4-gram, label unit=CDp2018.059.6
Povey et al.Architecture=HMM, Language Model=Yes2019.049.6
hybridLM=4-gram, label unit=CDp2018.059.8
MLP/CNN+I-Vector2014.1210.4
Soltau et al.AM training data=SWB2015.0510.4
LAS (Ours)Architecture=LAS, Language Model=Yes2019.0410.9
Zeyer et al. [38]Architecture=LAS, Language Model=Yes2019.0411
LAS (Ours)Architecture=LAS, Language Model=No2019.0411.2
CNN-HMM2014.1211.5
WDXNumber of convolutional layers=10, Optimization strategy=SGD from random initialization, # params (M)=41.3, #M frames=3202015.0911.8
attention modelLM=LSTM, label unit=BPE 1K2018.0511.8
Zeyer et al. [24]Architecture=LAS, Language Model=Yes2019.0411.8
VDXNumber of convolutional layers=8, Optimization strategy=SGD from random initialization, # params (M)=38.4, #M frames=3402015.0911.9
Zeyer et al. [38]Architecture=LAS, Language Model=No2019.0411.9
WDXNumber of convolutional layers=10, Optimization strategy=Adadelta + SGD finetuning, # params (M)=41.3, #M frames=1402015.0912.2
Weng et al.Architecture=LAS, Language Model=No2019.0412.2
VDXNumber of convolutional layers=8, Optimization strategy=Adadelta + SGD finetuning, # params (M)=38.4, #M frames=1702015.0912.3
DNN-HMM sMBR2014.1212.6
Deep SpeechTraining Data=SWB + FSH2014.1212.6
Vesely et al.AM training data=SWB2015.0512.6
Hannun et al.AM training data=SWB+Fisher2015.0512.6
Veselý et al.Architecture=HMM, Language Model=Yes2019.0412.9
Seide et al.AM training data=SWB+Fisher+other2015.0513.1
VCXNumber of convolutional layers=6, Optimization strategy=Adadelta + SGD finetuning, # params (M)=36.9, #M frames=2902015.0913.1
attention modelLM=none, label unit=BPE 1K2018.0513.1
Zeyer et al. [24]Architecture=LAS, Language Model=No2019.0413.1
Classic 512# params (M)=41.2, #M frames=12002015.0913.2
DNN-HMM sMBR HF2014.1213.3
attention modelLM=none, label unit=BPE 10K2018.0513.5
Classic 256 ReLUOptimization strategy=Adadelta + SGD finetuning, # params (M)=58.7, #M frames=2902015.0913.8
CTCLM=word RNN, label unit=chars, added lexicon=true2018.0514
Zweig et al.Architecture=CTC, Language Model=Yes2019.0414
LAS + SpecAugment (SS)Architecture=LAS, Policy=SS, Language Model=Yes2019.0414
LAS + SpecAugment (SM)Architecture=LAS, Policy=SM, Language Model=Yes2019.0414.1
Zhou et al.AM training data=SWB2015.0514.2
Maas et al.AM training data=SWB2015.0514.3
LAS + SpecAugment (SS)Architecture=LAS, Policy=SS, Language Model=No2019.0414.4
DNN-HMMTraining Data=SWB2014.1214.6
Audhkhasi et al. [45]Architecture=CTC, Language Model=No2019.0414.6
LAS + SpecAugment (SM)Architecture=LAS, Policy=SM, Language Model=No2019.0414.6
Maas et al.AM training data=SWB+Fisher2015.0515
DNN-HMMTraining Data=FSH2014.1216
CD-DNN2014.1216.1
Zeyer et al. [24]Architecture=HMM, Language Model=Yes2019.0417.3
GMM-HMM BMMI2014.1218.6
Hadian et al.Architecture=HMM, Language Model=Yes2019.0418.9
Povey et al.Architecture=HMM, Language Model=Yes2019.0419.3
LAS (Ours)Architecture=LAS, Language Model=Yes2019.0419.4
CTCLM=n-gram, label unit=chars2018.0519.8
Deep SpeechTraining Data=SWB2014.1220
CTCLM=RNN, label unit=chars, added noise=true2018.0520
Audhkhasi et al. [44]Architecture=CTC, Language Model=No2019.0420.8
CTCLM=RNN, label unit=chars2018.0521.4
LAS (Ours)Architecture=LAS, Language Model=No2019.0421.6
attentionLM=none, label unit=chars2018.0523.1
Toshniwal et al.Architecture=LAS, Language Model=No2019.0423.1
Zeyer et al. [38]Architecture=LAS, Language Model=Yes2019.0423.1
Weng et al.Architecture=LAS, Language Model=No2019.0423.3
Audhkhasi et al. [45]Architecture=CTC, Language Model=No2019.0423.6
Zeyer et al. [38]Architecture=LAS, Language Model=No2019.0423.7
Veselý et al.Architecture=HMM, Language Model=Yes2019.0424.5
CTCLM=none, label unit=chars2018.0524.7
Zweig et al.Architecture=CTC, Language Model=No2019.0424.7
Zweig et al.Architecture=CTC, Language Model=Yes2019.0425.3
Zeyer et al. [24]Architecture=LAS, Language Model=Yes2019.0425.7
attentionLM=3-gram, label unit=words2018.0525.8
Lu et al.Architecture=LAS, Language Model=Yes2019.0425.8
Zeyer et al. [24]Architecture=LAS, Language Model=No2019.0426.1
attentionLM=none, label unit=words2018.0526.8
Lu et al.Architecture=LAS, Language Model=No2019.0426.8
Audhkhasi et al. [44]Architecture=CTC, Language Model=No2019.0430.4
attentionLM=5-gram, label unit=chars2018.0530.5
attentionLM=none, label unit=chars2018.0532.8
Zweig et al.Architecture=CTC, Language Model=No2019.0437.1
CTCLM=none, label unit=chars2018.0538
Toshniwal et al.Architecture=LAS, Language Model=No2019.0440.8
Lu et al.Architecture=LAS, Language Model=Yes2019.0446
Lu et al.Architecture=LAS, Language Model=No2019.0448.2