List your company
ResearchFormerly WizWand

Speech Recognition on WSJ nov93 (dev)

3WERLF-MMI
2.2327.41612.617.784Oct 12, 2019Oct 24, 2020Nov 7, 2021Nov 21, 2022Dec 4, 2023Dec 17, 2024Dec 31, 2025
Updated 8mo ago

Evaluation results

MethodLinks
LF-MMIUnit=bi-phone, LM=4-gram2020.053——
Deep Speech 2Training=12K h labeled speech2019.104.42——
vq-wav2vec Gumbel + BERT baseLanguage Model=CHAR CONVLM2019.104.461.79—
Supervised transfer-learning2019.104.99——
wav2vecLanguage Model=CHAR CONVLM2019.105.12.11—
LAS Relaxed AttentionAED model type=LAS, # of AM par's=17.8M, gamma=0.102021.075.42—2.75
LAS Relaxed AttentionAED model type=LAS, # of AM par's=17.8M, gamma=0.152021.075.43—2.72
LAS Relaxed AttentionAED model type=LAS, # of AM par's=17.8M, gamma=0.052021.075.53—2.8
LAS Relaxed AttentionAED model type=LAS, # of AM par's=17.8M, gamma=0.202021.075.59—2.87
Lattice-free MMI2019.105.66——
CATUnit=mono-phone, LM=4-gram2020.055.7——
LAS BaselineAED model type=LAS, # of AM par's=17.8M, gamma=02021.075.7—2.97
Transformer Relaxed AttentionAED model type=Transformer, # of AM par's=16.8M, gamma=0.352021.075.8—3.24
Transformer Relaxed AttentionAED model type=Transformer, # of AM par's=16.8M, gamma=0.402021.075.83—3.19
LAS Relaxed AttentionAED model type=LAS, # of AM par's=17.8M, gamma=0.252021.075.99—3.14
LF-MMIUnit=mono-phone, LM=4-gram2020.056——
E2E-LF-MMIUnit=bi-phone, LM=4-gram2020.056——
LAS Relaxed AttentionAED model type=LAS, # of AM par's=17.8M, gamma=0.302021.076—3.02
Transformer Relaxed AttentionAED model type=Transformer, # of AM par's=16.8M, gamma=0.252021.076.02—3.32
Transformer Relaxed AttentionAED model type=Transformer, # of AM par's=16.8M, gamma=0.152021.076.09—3.54
Transformer Relaxed AttentionAED model type=Transformer, # of AM par's=16.8M, gamma=0.202021.076.14—3.45
vq-wav2vec Gumbel + BERT baseLanguage Model=4-gram LM2019.106.282.41—
E2E-LF-MMIUnit=mono-phone, LM=4-gram2020.056.3——
Transformer Relaxed AttentionAED model type=Transformer, # of AM par's=16.8M, gamma=0.302021.076.32—3.58
Transformer Relaxed AttentionAED model type=Transformer, # of AM par's=16.8M, gamma=0.052021.076.41—3.7
Transformer Relaxed AttentionAED model type=Transformer, # of AM par's=16.8M, gamma=0.102021.076.54—3.79
Baseline (log-mel)Language Model=CHAR CONVLM2019.106.672.77—
Transformer BaselineAED model type=Transformer, # of AM par's=16.8M, gamma=02021.076.69—3.91
Trainable frontend2019.106.8——
CTC/attentionUnit=mono-char, LM=RNN2020.056.8——
Moriya et al.2021.076.9——
wav2vecLanguage Model=4-gram LM2019.106.962.73—
LAS Relaxed AttentionAED model type=LAS, # of AM par's=17.8M, gamma=0.352021.077.1—3.81
Wav2letter++Unit=mono-char, LM=Conv2020.057.5——
CATUnit=mono-char, LM=4-gram2020.058.1——
Baseline (log-mel)Language Model=4-gram LM2019.108.573.32—
Wav2letter++Unit=mono-char, LM=4-gram2020.059.5——
vq-wav2vec GumbelLanguage Model=4-gram LM2019.109.553.93—
EESENUnit=mono-phone, LM=3-gram2020.0510.87——
ESPnetUnit=mono-char, LM=RNN2020.0512.4——
vq-wav2vec Gumbel + BERT baseLanguage Model=None2019.1013.44.13—
Masked-VPCProbing with seq2seq=true, Lexicon-free=true, No language model=true2025.1213.6—4.4
Masked-NCEProbing with seq2seq=true, Lexicon-free=true, No language model=true2025.1214.5—5
HuBERT ObjProbing with seq2seq=true, Lexicon-free=true, No language model=true2025.1215.2—5.2
Future-VPCProbing with seq2seq=true, Lexicon-free=true, No language model=true2025.1215.5—5.4
wav2vecLanguage Model=None2019.1016.245.07—
VQ-APCProbing with seq2seq=true, Lexicon-free=true, No language model=true2025.1216.8—5.8
log MelProbing with seq2seq=true, Lexicon-free=true, No language model=true2025.1218.2—6.8
Baseline (log-mel)Language Model=None2019.1019.466.28—
wav2letter++Probing with seq2seq=false, Lexicon-free=true, No language model=true2025.1219.5—6.3
vq-wav2vec GumbelLanguage Model=None2019.1020.447.04—
18L TransformerProbing with seq2seq=false, Lexicon-free=true, No language model=true2025.1222.2——