List your company
ResearchFormerly WizWand

Audio-Visual Speech Recognition on LRS3 Clean original (test)

0.65WERM2S-AVSR
0.52521.36762.213.0524Jun 4, 2026
Updated 5mo ago

Evaluation results

MethodLinks
M2S-AVSRModality=A+V, Total Params=2.6B, Noise Dataset=LRS3, Labeled Training Data (hrs)=ms1,759, Unlabeled Training Data (hrs)=ms1,759, Multi-view Synthesis=true2026.060.65
M2S-AVSRModality=A+V, Total Params=2.6B, Noise Dataset=LRS3, Labeled Training Data (hrs)=1,759, Unlabeled Training Data (hrs)=1,759, Multi-view Synthesis=false2026.060.68
MMS-LlamaModality=A+V, Total Params=3.2B, Noise Dataset=NoiseX, Labeled Training Data (hrs)=1,759, Unlabeled Training Data (hrs)=1,7592026.060.72
Llama-AVSRModality=A+V, Total Params=8.7B, Noise Dataset=NoiseX, Labeled Training Data (hrs)=1,759, Unlabeled Training Data (hrs)=1,7592026.060.77
M2S-AVSRModality=A+V, Total Params=2.6B, Noise Dataset=LRS3, Labeled Training Data (hrs)=433, Unlabeled Training Data (hrs)=1,759, Multi-view Synthesis=false2026.060.82
M2S-AVSRModality=A+V, Total Params=2.6B, Noise Dataset=LRS3, Labeled Training Data (hrs)=ms433, Unlabeled Training Data (hrs)=ms1,759, Multi-view Synthesis=true2026.060.82
Auto-AVSRModality=A+V, Total Params=443M, Noise Dataset=NoiseX, Labeled Training Data (hrs)=3,4482026.060.9
MMS-LlamaModality=A+V, Total Params=3.2B, Noise Dataset=NoiseX, Labeled Training Data (hrs)=433, Unlabeled Training Data (hrs)=1,7592026.060.9
Qwen3-ASRModality=A, Total Params=1.7B, Noise Dataset=LRS3, Unlabeled Training Data (hrs)=∼ 4.0 × 10^7, LLM-based (No Fine-tuning)=true2026.061.39
AV-HuBERTModality=A+V, Total Params=477M, Noise Dataset=LRS3, Labeled Training Data (hrs)=433, Unlabeled Training Data (hrs)=1,7592026.061.4
CMAModality=A+V, Total Params=505M, Noise Dataset=LRS3, Labeled Training Data (hrs)=433, Unlabeled Training Data (hrs)=1,7592026.061.5
Whisper-FlamingoModality=A+V, Total Params=2.5B, Noise Dataset=LRS3, Labeled Training Data (hrs)=433, Unlabeled Training Data (hrs)=1,7592026.061.5
Whisper-FlamingoModality=A+V, Total Params=2.5B, Noise Dataset=LRS3, Labeled Training Data (hrs)=ms433, Unlabeled Training Data (hrs)=ms1,759, Multi-view Synthesis=true2026.061.5
Whisper-FlamingoModality=A+V, Total Params=2.5B, Noise Dataset=LRS3, Labeled Training Data (hrs)=ms1,759, Unlabeled Training Data (hrs)=ms1,759, Multi-view Synthesis=true2026.061.95
Whisper-finetunedModality=A, Total Params=1.6B, Noise Dataset=LRS3, Labeled Training Data (hrs)=1,7592026.062
Whisper-FlamingoModality=A+V, Total Params=2.5B, Noise Dataset=LRS3, Labeled Training Data (hrs)=1,759, Unlabeled Training Data (hrs)=1,7592026.062
Whisper-finetunedModality=A, Total Params=1.6B, Noise Dataset=LRS3, Labeled Training Data (hrs)=4332026.062.3
AV-RelScoreModality=A+V, Total Params=136M, Noise Dataset=NoiseX, Labeled Training Data (hrs)=4332026.062.77
V-CAFEModality=A+V, Total Params=109M, Noise Dataset=NoiseX, Labeled Training Data (hrs)=4332026.062.82
Fun-ASR-NanoModality=A, Total Params=0.8B, Noise Dataset=LRS3, Labeled Training Data (hrs)=millions, Unlabeled Training Data (hrs)=tens of millions, LLM-based (No Fine-tuning)=true2026.062.96
FireRed-ASRModality=A, Total Params=1.1B, Noise Dataset=LRS3, Labeled Training Data (hrs)=∼ 8.1 × 10^4, LLM-based (No Fine-tuning)=true2026.063.77