List your company
ResearchFormerly WizWand

Automatic Speech Recognition on LibriSpeech 960h (test-clean)

0.016WERSPEARs+a XLarge
0.0148440.0226470.030450.038253May 19, 2020May 22, 2021May 26, 2022May 30, 2023Jun 1, 2024Jun 5, 2025Jun 9, 2026
Updated 4mo ago

Evaluation results

MethodLinks
SPEARs+a XLarge# Params=600M, Pre-train data=197k2025.100.016
Noisy Student TrainingBackbone=ContextNet, LM Fusion Stage=before2020.050.017
Noisy Student TrainingBackbone=ContextNet, LM Fusion Stage=after2020.050.017
SPEARs Large# Params=327M, Pre-train data=84k2025.100.017
SPEARs+a Large# Params=327M, Pre-train data=97k2025.100.017
wav2vec 2.0 LargeUnlabeled Data=LL-60k, LM=Transformer2021.100.018
WavLM LargeUnlabeled Data=MIX-94k, LM=Transformer2021.100.018
WavLM Large# Params=317M, Pre-train data=94k, fine-tuned from public checkpoint=true2025.100.018
ContextNetTraining Paradigm=Supervised2020.050.019
BaselineLanguage Model=Transformer2021.100.019
SpeechT5Language Model=Transformer2021.100.019
ContextNetUnlabeled Data=N/A, LM=LSTM2021.100.019
Conformer TransducerUnlabeled Data=N/A, LM=LSTM2021.100.019
HuBERT LargeUnlabeled Data=LL-60k, LM=Transformer2021.100.019
SPEARs Base# Params=94M, Pre-train data=84k2025.100.019
SPEARs+a Base# Params=94M, Pre-train data=97k2025.100.019
Zhang et al.Training Paradigm=Supervised2020.050.02
Transf. TransducerUnlabeled Data=N/A, LM=Transf.2021.100.02
HuBERT-LTrain=CTC, Decode=Greedy2026.060.02
HuBERT-L + Llama 8B (4bit, r=32)Train=Joint, Decode=S2T2026.060.02
Synnaeve et al.Training Paradigm=Semi-supervised2020.050.0209
wav2vec 2.0 BASELanguage Model=Transformer2021.100.021
Branchformer + GFSA#Params=109.8M2023.120.0211
Branchformer#Params=109.8M2023.120.0213
Whisper-L + Llama 8B (4bit, r=32)Train=Joint, Decode=S2T2026.060.022
HuBERT-L + Llama 8B (4bit, r=32)Train=Joint, Decode=PF-S2T2026.060.022
S2S Transf.Unlabeled Data=N/A, LM=CLM+Transf.2021.100.023
Whisper-LDecode=Beam2026.060.023
Transformer + GFSA#Params=71.5M2023.120.0231
Synnaeve et al.Training Paradigm=Supervised2020.050.0233
Transformer#Params=71.5M2023.120.0242
CTC Transf.Unlabeled Data=N/A, LM=CLM+Transf.2021.100.025
wav2vec 2.0 BASELanguage Model=4-gram2021.100.026
Whisper-L + Llama 8B (4bit, r=32)Train=Joint, Decode=PF-S2T2026.060.026
HuBERT-L + Llama 8B (4bit, r=32)Train=PF-S2T, Decode=PF-S2T2026.060.026
SpeechT5Language Model=None2021.100.027
PDS-Base-16Architecture=Conformer, Group=E, L=12, dh=512, dff=2048, h=8, #Params=114M, Speedup=1.23x2023.050.0272
BaselineLanguage Model=None2021.100.028
Whisper-L + Llama 8B (4bit, r=32)Train=PF-S2T, Decode=PF-S2T2026.060.028
HuBERT-L + Llama 8B (4bit, r=32)Train=S2T, Decode=S2T2026.060.028
Stack-4 (Baseline)Architecture=Conformer, Group=E, L=12, dh=512, dff=2048, h=8, #Params=113M, Speedup=1.00x2023.050.0282
WavLM Base +# Params=95M, Pre-train data=94k, fine-tuned from public checkpoint=true2025.100.029
PDS-Base-32Architecture=Conformer, Group=E, L=12, dh=512, dff=2048, h=8, #Params=119M, Speedup=1.25x2023.050.0295
PDS-Base-16Architecture=Conformer, Group=D, L=12, dh=256, dff=2048, h=4, #Params=46M, Speedup=1.14x2023.050.0299
Stack-4 (wenet)Architecture=Conformer, Group=E, L=12, dh=512, dff=2048, h=8, #Params=109M2023.050.03
PDS-Base-8Architecture=Conformer, Group=D, L=12, dh=256, dff=2048, h=4, #Params=46M, Speedup=0.97x2023.050.0301
AIPAArchitecture=Enc-CTC, Data Augmentation=SpecAugment + AIPA, Auxiliary Loss=InterCTC, Technique=COS2024.060.0301
PDS-Base-8Architecture=Conformer, Group=E, L=12, dh=512, dff=2048, h=8, #Params=113M, Speedup=0.97x2023.050.0303
PDS-Base-32Architecture=Conformer, Group=D, L=12, dh=256, dff=2048, h=4, #Params=47M, Speedup=1.20x2023.050.0304
Stack-4 (wenet)Architecture=Conformer, Group=D, L=12, dh=256, dff=2048, h=4, #Params=50M2023.050.0305
Baseline (w/o CTC)Language Model=None2021.100.031
SpeechT5 (w/o CTC)Language Model=None2021.100.031
Stack-4 (Baseline)Architecture=Conformer, Group=D, L=12, dh=256, dff=2048, h=4, #Params=45M, Speedup=1.00x2023.050.0321
Whisper-L + Llama 8B (4bit, r=32)Train=S2T, Decode=S2T2026.060.033
PDS-Deep-32Architecture=Transformer, Group=C, L=30, dh=256, dff=2048, h=4, #Params=55M, Speedup=1.27x2023.050.0333
PDS-Base-16Architecture=Transformer, Group=B, L=12, dh=512, dff=2048, h=8, #Params=76M, Speedup=1.34x2023.050.0337
PDS-Deep-16Architecture=Transformer, Group=C, L=30, dh=256, dff=2048, h=4, #Params=54M, Speedup=1.19x2023.050.0338
wav2vec 2.0 BASELanguage Model=None2021.100.034
Stack-4 (fairseq)Architecture=Transformer, Group=B, L=12, dh=512, dff=2048, h=8, #Params=71M2023.050.034
PDS-Base-8Architecture=Transformer, Group=B, L=12, dh=512, dff=2048, h=8, #Params=75M, Speedup=1.08x2023.050.0347
PDS-Deep-8Architecture=Transformer, Group=C, L=30, dh=256, dff=2048, h=4, #Params=53M, Speedup=1.03x2023.050.035
BaselineArchitecture=Enc-CTC, Data Augmentation=SpecAugment, Auxiliary Loss=InterCTC2024.060.0361
PDS-Base-32Architecture=Transformer, Group=B, L=12, dh=512, dff=2048, h=8, #Params=82M, Speedup=1.47x2023.050.0364
Stack-4 (Baseline)Architecture=Transformer, Group=B, L=12, dh=512, dff=2048, h=8, #Params=71M, Speedup=1.00x2023.050.0367
PDS-Base-16Architecture=Transformer, Group=A, L=12, dh=256, dff=2048, h=4, #Params=30M, Speedup=1.14x2023.050.0374
PDS-Base-8Architecture=Transformer, Group=A, L=12, dh=256, dff=2048, h=4, #Params=30M, Speedup=0.99x2023.050.0385
PDS-Base-32Architecture=Transformer, Group=A, L=12, dh=256, dff=2048, h=4, #Params=31M, Speedup=1.20x2023.050.0421
Stack-4 (Baseline)Architecture=Transformer, Group=C, L=30, dh=256, dff=2048, h=4, #Params=53M, Speedup=1.00x2023.050.0433
Stack-4 (fairseq)Architecture=Transformer, Group=A, L=12, dh=256, dff=2048, h=4, #Params=30M2023.050.044
Stack-4 (Baseline)Architecture=Transformer, Group=A, L=12, dh=256, dff=2048, h=4, #Params=30M, Speedup=1.00x2023.050.0449