List your company
ResearchFormerly WizWand

Automatic Speech Recognition on AISHELL-1 (dev)

0.43CERNIM4-ASR
0.20281.73643.274.8036Jan 30, 2023Aug 14, 2023Feb 26, 2024Sep 9, 2024Mar 24, 2025Oct 6, 2025Apr 20, 2026
Updated 4mo ago

Evaluation results

MethodLinks
NIM4-ASRModel Size=2.3B, Inference Mode=Offline2026.040.43
NIM4-ASRModel Size=2.3B, Inference Mode=Stream2026.040.43
OursModel Size=2.3B2026.040.45
FireRedASR2S LLMModel Size=8B+, Inference Mode=Offline2026.040.6
FireRedASR-LLMModel Size=8B+ (↑)2026.040.71
Step-Audio2-miniModel Size=8B+ (↑)2026.040.76
Step-Audio2 MiniModel Size=8B+, Inference Mode=Offline2026.040.76
Qwen3-Omni-InstModel Size=30B-A3B (↑)2026.040.86
Qwen3-Omni InstructModel Size=30B-A3B, Inference Mode=Offline2026.040.86
Qwen3-ASR-1.7BModel Size=2.0B (↓)2026.041.4
Qwen3-ASR 1.7BModel Size=2.0B, Inference Mode=Offline2026.041.4
Fun-ASR-nanoModel Size=0.8B (↓)2026.041.59
Fun-ASR NanoModel Size=0.8B, Inference Mode=Offline2026.041.59
GLM-ASR-nanoModel Size=1.5B (↓)2026.042.4
GLM-ASR NanoModel Size=1.5B, Inference Mode=Offline2026.042.4
UOT-BERT-CTClambda1=0.5, lambda2=1.0, epsilon=0.05, eta=0.3, Decoding strategy=CTC greedy search-based2025.093.64
UOT-BERT-CTClambda1=1.0, lambda2=1.0, epsilon=0.05, eta=0.3, Decoding strategy=CTC greedy search-based2025.093.7
Neural TransducerLM=true, # Param=90 M2023.013.8
CIF + HKDLM=false, # Param=47 M2023.013.8
CIF + HKDLM=true, # Param=47 M2023.013.8
OT-BERT-CTCDecoding strategy=CTC greedy search-based, eta=0.32025.093.81
UOT-BERT-CTClambda1=1.0, lambda2=0.5, epsilon=0.05, eta=0.3, Decoding strategy=CTC greedy search-based2025.093.81
UOT-BERT-CTClambda1=10.0, lambda2=10.0, epsilon=0.05, eta=0.3, Decoding strategy=CTC greedy search-based2025.093.82
CIF + LRDLM=false, # Param=47 M2023.014
CIF + LRDLM=true, # Param=47 M2023.014
Zipformer-LParams (M)=157.3, Type=pruned transducer2023.104.03
Zipformer-MParams (M)=73.4, Type=pruned transducer2023.104.13
E-Branchformer, hybridType=AR, Params=57M, LM usage=without LM, Decoding beam size=102025.094.13
UMA-Split (prop.)Type=NAR, Params=46M, LM usage=without LM2025.094.15
NAR-BERT-ASRDecoding strategy=CTC greedy search-based, eta=0.32025.094.18
BranchformerParams (M)=45.4, Type=CTC/AED2023.104.19
Branchformer (B), hybridType=AR, Params=45M, LM usage=without LM, Decoding beam size=102025.094.19
BranchformerLM=false, # Param=45 M2023.014.2
CIF + ACDLM=false, # Param=47 M2023.014.2
CIF + ACDLM=true, # Param=47 M2023.014.2
E-Branchformer in ESPnetParams (M)=37.9, Type=CTC/AED2023.104.2
LAStraining hours=18.12024.094.26
EffectiveASR LargeType=NAR, Params=76M, LM usage=without LM2025.094.26
UMA-Split (w/o split module)Type=NAR, Params=45M, LM usage=without LM2025.094.28
Hybrid (Conformer)Search=beam search, RTF=0.461, Params (M)=46.32023.094.3
Conformer+CTC/AEDDecoding strategy=CTC greedy search-based, eta=0.32025.094.31
CTC Infer w/o AED headType=NAR, Params=46M, LM usage=without LM2025.094.39
ESPnet ConformerLM=true, # Param=46 M2023.014.4
CIFLM=true, # Param=47 M2023.014.4
UMA + self-conditionRTF=0.045, Params (M)=44.72023.094.4
Zipformer-SParams (M)=30.2, Type=pruned transducer2023.104.4
Original UMA ConformerType=NAR, Params=45M, LM usage=without LM2025.094.4
Lightweight Transducertraining hours=18.62024.094.45
Zipformer-M, CTCType=NAR, Params=66M, LM usage=without LM2025.094.47
ESPnet ConformerLM=false, # Param=46 M2023.014.5
CIFLM=false, # Param=47 M2023.014.5
UMARTF=0.039, Params (M)=42.62023.094.5
Conformer in ESPnetParams (M)=46.2, Type=CTC/AED2023.104.5
Paraformer-v2 (S)Type=NAR, Params=50M, LM usage=without LM2025.094.5
Transducertraining hours=64.22024.094.59
Paraformer2023.094.6
Self-conditioned CTCRTF=0.059, Params (M)=51.52023.094.6
Hybrid (Conformer)RTF=0.125, Params (M)=46.32023.095
Conformer+CTCDecoding strategy=CTC greedy search-based, eta=0.32025.095.16
CTCRTF=0.052, Params (M)=50.42023.095.6
RNN-TRd=-, Ru=-2023.055.8
BATRd=2, Ru=22023.055.88
BATRd=1, Ru=12023.055.89
LASO-largeParams (M)=80.02023.095.9
RNN-T + FElambda=0.004, Rd=-, Ru=-2023.056.04
RNN-T + FElambda=0.002, Rd=-, Ru=-2023.056.11