List your company
ResearchFormerly WizWand

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

2.5WERw2v-BERT XXL
2.22324.09165.967.8284Aug 7, 2021May 28, 2022Mar 19, 2023Jan 7, 2024Oct 28, 2024Aug 18, 2025Jun 9, 2026
Updated 4mo ago

Evaluation results

MethodLinks
w2v-BERT XXLUnlabeled Data (hrs)=60k, AM Size (B)=1.0, LM Size (B)=0.1, Training Strategy=Pre-training + Self-training, Decoding Strategy=No LM2021.082.5——
w2v-BERT XXLUnlabeled Data (hrs)=60k, AM Size (B)=1.0, LM Size (B)=0.1, Training Strategy=Pre-training + Self-training, Decoding Strategy=With LM2021.082.5——
w2v-Conformer XXL+Unlabeled Data (hrs)=60k, AM Size (B)=1.1, LM Size (B)=0.1, Training Strategy=Pre-training + Self-training, Decoding Strategy=With LM2021.082.6——
w2v-BERT XLUnlabeled Data (hrs)=60k, AM Size (B)=0.6, LM Size (B)=0.1, Training Strategy=Pre-training + Self-training, Decoding Strategy=With LM2021.082.6——
w2v-BERT XXLUnlabeled Data (hrs)=60k, AM Size (B)=1.0, LM Size (B)=0.1, Training Strategy=Pre-training Only, Decoding Strategy=With LM2021.082.7——
w2v-Conformer XXLUnlabeled Data (hrs)=60k, AM Size (B)=1.0, LM Size (B)=0.1, Training Strategy=Pre-training + Self-training, Decoding Strategy=With LM2021.082.7——
w2v-Conformer XXL+Unlabeled Data (hrs)=60k, AM Size (B)=1.1, LM Size (B)=0.1, Training Strategy=Pre-training + Self-training, Decoding Strategy=No LM2021.082.7——
w2v-BERT XLUnlabeled Data (hrs)=60k, AM Size (B)=0.6, LM Size (B)=0.1, Training Strategy=Pre-training + Self-training, Decoding Strategy=No LM2021.082.7——
w2v-BERT XLUnlabeled Data (hrs)=60k, AM Size (B)=0.6, LM Size (B)=0.1, Training Strategy=Pre-training Only, Decoding Strategy=With LM2021.082.8——
w2v-BERT XXLUnlabeled Data (hrs)=60k, AM Size (B)=1.0, LM Size (B)=0.1, Training Strategy=Pre-training Only, Decoding Strategy=No LM2021.082.8——
w2v-Conformer XXLUnlabeled Data (hrs)=60k, AM Size (B)=1.0, LM Size (B)=0.1, Training Strategy=Pre-training + Self-training, Decoding Strategy=No LM2021.082.8——
HuBERT X-LargeUnlabeled Data (hrs)=60k, AM Size (B)=1.0, Training Strategy=Pre-training Only, Decoding Strategy=With LM2021.082.9——
w2v-BERT XLUnlabeled Data (hrs)=60k, AM Size (B)=0.6, LM Size (B)=0.1, Training Strategy=Pre-training Only, Decoding Strategy=No LM2021.082.9——
SPEARs+a XLarge# Params=600M, Pre-train data=197k2025.102.9——
w2v-Conformer XXLUnlabeled Data (hrs)=60k, AM Size (B)=1.0, LM Size (B)=0.1, Training Strategy=Pre-training Only, Decoding Strategy=With LM2021.083.1——
wav2vec 2.0Unlabeled Data (hrs)=60k, AM Size (B)=0.3, LM Size (B)=> 0.4, Training Strategy=Pre-training + Self-training, Decoding Strategy=With LM2021.083.1——
WavLM LargeUnlabeled Data=MIX-94k, LM=Transformer2021.103.2——
w2v-Conformer XLUnlabeled Data (hrs)=60k, AM Size (B)=0.6, LM Size (B)=0.1, Training Strategy=Pre-training Only, Decoding Strategy=With LM2021.083.2——
wav2vec 2.0 LargeUnlabeled Data=LL-60k, LM=Transformer2021.103.3——
HuBERT LargeUnlabeled Data=LL-60k, LM=Transformer2021.103.3——
Conformer L with NSTUnlabeled Data (hrs)=60k, AM Size (B)=0.1, LM Size (B)=0.1, Training Strategy=Self-training Only, Decoding Strategy=With LM2021.083.3——
wav2vec 2.0Unlabeled Data (hrs)=60k, AM Size (B)=0.3, LM Size (B)=> 0.4+, Training Strategy=Pre-training Only, Decoding Strategy=With LM2021.083.3——
HuBERT LargeUnlabeled Data (hrs)=60k, AM Size (B)=0.3, Training Strategy=Pre-training Only, Decoding Strategy=With LM2021.083.3——
w2v-Conformer XXLUnlabeled Data (hrs)=60k, AM Size (B)=1.0, LM Size (B)=0.1, Training Strategy=Pre-training Only, Decoding Strategy=No LM2021.083.3——
SPEARs Large# Params=327M, Pre-train data=84k2025.103.3——
SPEARs+a Large# Params=327M, Pre-train data=97k2025.103.4——
Conformer L with NSTUnlabeled Data (hrs)=60k, AM Size (B)=0.1, LM Size (B)=0.1, Training Strategy=Self-training Only, Decoding Strategy=No LM2021.083.5——
w2v-Conformer XLUnlabeled Data (hrs)=60k, AM Size (B)=0.6, LM Size (B)=0.1, Training Strategy=Pre-training Only, Decoding Strategy=No LM2021.083.5——
wav2vec 2.0Unlabeled Data (hrs)=60k, AM Size (B)=0.3, LM Size (B)=> 0.4, Training Strategy=Pre-training + Self-training, Decoding Strategy=No LM2021.083.5——
WavLM Large# Params=317M, Pre-train data=94k, fine-tuned from public checkpoint=true2025.103.8——
Conformer TransducerUnlabeled Data=N/A, LM=LSTM2021.103.9——
Conformer LUnlabeled Data (hrs)=N/A, AM Size (B)=0.1, LM Size (B)=0.1, Training Strategy=Trained from Scratch, Decoding Strategy=With LM2021.083.9——
SPEARs Base# Params=94M, Pre-train data=84k2025.104——
ContextNetUnlabeled Data=N/A, LM=LSTM2021.104.1——
Whisper-LDecode=Beam2026.064.1——
HuBERT-LTrain=CTC, Decode=Greedy2026.064.1——
SPEARs+a Base# Params=94M, Pre-train data=97k2025.104.2——
Conformer LUnlabeled Data (hrs)=N/A, AM Size (B)=0.1, LM Size (B)=0.1, Training Strategy=Trained from Scratch, Decoding Strategy=No LM2021.084.3——
HuBERT-L + Llama 8B (4bit, r=32)Train=Joint, Decode=S2T2026.064.4——
wav2vec 2.0Unlabeled Data (hrs)=60k, AM Size (B)=0.3, LM Size (B)=> 0.4+, Training Strategy=Pre-training Only, Decoding Strategy=No LM2021.084.5——
Transf. TransducerUnlabeled Data=N/A, LM=Transf.2021.104.6——
HuBERT-L + Llama 8B (4bit, r=32)Train=Joint, Decode=PF-S2T2026.064.9——
Branchformer + GFSA#Params=109.8M2023.124.94——
Branchformer#Params=109.8M2023.125——
Whisper-L + Llama 8B (4bit, r=32)Train=Joint, Decode=S2T2026.065.1——
S2S Transf.Unlabeled Data=N/A, LM=CLM+Transf.2021.105.2——
HuBERT-L + Llama 8B (4bit, r=32)Train=S2T, Decode=S2T2026.065.3——
WavLM Base +# Params=95M, Pre-train data=94k, fine-tuned from public checkpoint=true2025.105.4——
Transformer + GFSA#Params=71.5M2023.125.49——
CTC Transf.Unlabeled Data=N/A, LM=CLM+Transf.2021.105.5——
Transformer#Params=71.5M2023.125.5——
Whisper-L + Llama 8B (4bit, r=32)Train=S2T, Decode=S2T2026.065.6——
HuBERT-L + Llama 8B (4bit, r=32)Train=PF-S2T, Decode=PF-S2T2026.065.6——
Proposed MethodBias list size (N)=200, Biasing weight (lambda)=4.42025.125.6317.165.22
Proposed MethodBias list size (N)=100, Biasing weight (lambda)=4.42025.125.6417.225.22
Proposed MethodBias list size (N)=500, Biasing weight (lambda)=4.42025.125.6417.555.21
Proposed MethodBias list size (N)=100, Biasing weight (lambda)=1.02025.125.8221.855.24
Proposed MethodBias list size (N)=200, Biasing weight (lambda)=1.02025.125.8221.855.24
Proposed MethodBias list size (N)=500, Biasing weight (lambda)=1.02025.125.8822.685.27
Whisper-L + Llama 8B (4bit, r=32)Train=Joint, Decode=PF-S2T2026.065.9——
AED-MTPBias list size (N)=02025.12630.635.12
AED-MTPBias list size (N)=1002025.12630.635.12
AED-MTPBias list size (N)=2002025.12630.635.12
AED-MTPBias list size (N)=5002025.12630.635.12
Baseline (AED)Bias list size (N)=02025.126.0132.345.07
Whisper-L + Llama 8B (4bit, r=32)Train=PF-S2T, Decode=PF-S2T2026.066.1——
Proposed MethodBias list size (N)=0, Biasing weight (lambda)=1.02025.126.2532.45.32
Proposed MethodBias list size (N)=0, Biasing weight (lambda)=4.42025.126.2532.45.32
PDS-Base-8Architecture=Conformer, Group=E, L=12, dh=512, dff=2048, h=8, #Params=113M2023.056.38——
PDS-Base-16Architecture=Conformer, Group=E, L=12, dh=512, dff=2048, h=8, #Params=114M2023.056.4——
Stack-4 (wenet)Architecture=Conformer, Group=E, L=12, dh=512, dff=2048, h=8, #Params=109M2023.056.7——
Stack-4 (Baseline)Architecture=Conformer, Group=E, L=12, dh=512, dff=2048, h=8, #Params=113M2023.056.72——
PDS-Base-32Architecture=Conformer, Group=E, L=12, dh=512, dff=2048, h=8, #Params=119M2023.056.81——
CLASBias list size (N)=1002025.126.8927.816.14
CLASBias list size (N)=2002025.126.8927.596.15
CLASBias list size (N)=5002025.126.9327.926.18
CLASBias list size (N)=02025.126.9528.266.19
PDS-Base-16Architecture=Conformer, Group=D, L=12, dh=256, dff=2048, h=4, #Params=46M2023.057.03——
PDS-Base-32Architecture=Conformer, Group=D, L=12, dh=256, dff=2048, h=4, #Params=47M2023.057.12——
PDS-Base-8Architecture=Conformer, Group=D, L=12, dh=256, dff=2048, h=4, #Params=46M2023.057.2——
Stack-4 (Baseline)Architecture=Conformer, Group=D, L=12, dh=256, dff=2048, h=4, #Params=45M2023.057.29——
PDS-Base-8Architecture=Transformer, Group=B, L=12, dh=512, dff=2048, h=8, #Params=75M2023.057.47——
AIPAArchitecture=Enc-CTC, Data Augmentation=SpecAugment + AIPA, Auxiliary Loss=InterCTC, Technique=COS2024.067.51——
PDS-Deep-8Architecture=Transformer, Group=C, L=30, dh=256, dff=2048, h=4, #Params=53M2023.057.79——
PDS-Deep-16Architecture=Transformer, Group=C, L=30, dh=256, dff=2048, h=4, #Params=54M2023.057.79——
PDS-Base-16Architecture=Transformer, Group=B, L=12, dh=512, dff=2048, h=8, #Params=76M2023.057.85——
PDS-Base-32Architecture=Transformer, Group=B, L=12, dh=512, dff=2048, h=8, #Params=82M2023.057.85——
PDS-Deep-32Architecture=Transformer, Group=C, L=30, dh=256, dff=2048, h=4, #Params=55M2023.057.88——
Stack-4 (fairseq)Architecture=Transformer, Group=B, L=12, dh=512, dff=2048, h=8, #Params=71M2023.057.9——
Stack-4 (Baseline)Architecture=Transformer, Group=B, L=12, dh=512, dff=2048, h=8, #Params=71M2023.057.96——
Stack-4 (wenet)Architecture=Conformer, Group=D, L=12, dh=256, dff=2048, h=4, #Params=50M2023.058.36——
PDS-Base-8Architecture=Transformer, Group=A, L=12, dh=256, dff=2048, h=4, #Params=30M2023.058.58——
Stack-4 (Baseline)Architecture=Transformer, Group=C, L=30, dh=256, dff=2048, h=4, #Params=53M2023.058.61——
Stack-4 (fairseq)Architecture=Transformer, Group=A, L=12, dh=256, dff=2048, h=4, #Params=30M2023.059——
PDS-Base-16Architecture=Transformer, Group=A, L=12, dh=256, dff=2048, h=4, #Params=30M2023.059.02——
BaselineArchitecture=Enc-CTC, Data Augmentation=SpecAugment, Auxiliary Loss=InterCTC2024.069.02——
PDS-Base-32Architecture=Transformer, Group=A, L=12, dh=256, dff=2048, h=4, #Params=31M2023.059.31——
Stack-4 (Baseline)Architecture=Transformer, Group=A, L=12, dh=256, dff=2048, h=4, #Params=30M2023.059.42——