List your company
ResearchFormerly WizWand

Speech Emotion Recognition on IEMOCAP (five-fold/ten-fold cross-validation)

77.64WAemotion2vec
54.312860.368966.42572.4811Dec 23, 2023May 10, 2024Sep 27, 2024Feb 13, 2025Jul 3, 2025Nov 19, 2025Apr 8, 2026
Updated 6mo ago

Evaluation results

MethodLinks
emotion2vecModel Scale=base size, Pre-training Corpus=LS-960 + Emo-262, #Upstream Params=93.79M, Downstream=Linear, #Downstream Params=0.20M, Evaluation Protocol=ten-fold leave-one-speaker-out, Same validation and test fold=true2023.1277.64
emotion2vecModel Scale=base size, Pre-training Corpus=LS-960 + Emo-262, #Upstream Params=93.79M, Downstream=Linear, #Downstream Params=0.20M, Evaluation Protocol=five-fold leave-one-session-out, Same validation and test fold=true2023.1274.48
emotion2vecModel Scale=base size, Pre-training Corpus=LS-960 + Emo-262, #Upstream Params=93.79M, Downstream=Linear, #Downstream Params=0.20M, Evaluation Protocol=ten-fold leave-one-speaker-out, Same validation and test fold=false2023.1272.94
DSTModel Scale=Supervised Model, Upstream=WavLM-large, #Upstream Params=316.62M, Downstream=Transformer(DST), #Downstream Params=22.78M2023.1271.8
emotion2vecModel Scale=base size, Pre-training Corpus=LS-960 + Emo-262, #Upstream Params=93.79M, Downstream=Linear, #Downstream Params=0.20M, Evaluation Protocol=five-fold leave-one-session-out, Same validation and test fold=false2023.1271.79
Vesper-12Model Scale=base size, Pre-training Corpus=Mix-94k + LSED-206, #Upstream Params=164.29M, Downstream=Linear, #Downstream Params=0.26M, Evaluation Protocol=five-fold leave-one-session-out2023.1270.7
WavLMModel Scale=large size, Pre-training Corpus=Mix-94k, #Upstream Params=316.62M, Downstream=Linear, Evaluation Protocol=five-fold leave-one-session-out2023.1270.03
MSTRModel Scale=Supervised Model, Upstream=HuBERT-large, #Upstream Params=316.61M, Downstream=Transformer(MSTR), #Downstream Params=27.00M2023.1270.03
data2vec 2.0Model Scale=base size, Pre-training Corpus=LS-960, #Upstream Params=93.78M, Downstream=Linear, #Downstream Params=0.20M, Evaluation Protocol=five-fold leave-one-session-out2023.1268.58
Vesper-4Model Scale=base size, Pre-training Corpus=Mix-94k + LSED-206, Upstream=Proposed, #Upstream Params=63.52M, Downstream=Linear, #Downstream Params=0.26M, Evaluation Protocol=five-fold leave-one-session-out2023.1268.4
HuBERT#Param. (M)=95, Data (hrs)=960, Evaluation Protocol=Fine-tuning2026.0468.3
TIM-NetModel Scale=Supervised Model, Upstream=MFCC, Downstream=CNN(TIM-Net), #Downstream Params=0.40M2023.1268.29
WavLM+Model Scale=base size, Pre-training Corpus=Mix-94k, #Upstream Params=94.70M, Downstream=Linear, #Downstream Params=0.20M, Evaluation Protocol=five-fold leave-one-session-out2023.1267.98
EnCodecMAE#Param. (M)=86.6, Data (hrs)=11,000, Evaluation Protocol=Fine-tuning2026.0467.8
ULTRAS#Param. (M)=89, Data (hrs)=2,000, Evaluation Protocol=Fine-tuning2026.0467.78
HUBERTModel Scale=large size, Pre-training Corpus=LL-60k, Upstream=Proposed, #Upstream Params=316.61M, Downstream=Linear, #Downstream Params=0.26M, Evaluation Protocol=five-fold leave-one-session-out2023.1267.62
BEATs#Param. (M)=91, Data (hrs)=5,800, Evaluation Protocol=Fine-tuning2026.0467.61
data2vecModel Scale=base size, Pre-training Corpus=LS-960, #Upstream Params=93.75M, Downstream=Linear, #Downstream Params=0.20M, Evaluation Protocol=five-fold leave-one-session-out2023.1267.38
WavLMModel Scale=base size, Pre-training Corpus=LS-960, #Upstream Params=94.70M, Downstream=Linear, #Downstream Params=0.20M, Evaluation Protocol=five-fold leave-one-session-out2023.1265.94
wav2vec 2.0Model Scale=large size, Pre-training Corpus=LL-60k, #Upstream Params=317.38M, Downstream=Linear, Evaluation Protocol=five-fold leave-one-session-out2023.1265.64
HUBERTModel Scale=base size, Pre-training Corpus=LS-960, #Upstream Params=94.68M, Downstream=Linear, #Downstream Params=0.20M, Evaluation Protocol=five-fold leave-one-session-out2023.1264.92
wav2vec 2.0Model Scale=base size, Pre-training Corpus=LS-960, #Upstream Params=95.04M, Downstream=Linear, #Downstream Params=0.20M, Evaluation Protocol=five-fold leave-one-session-out2023.1263.43
wav2vecModel Scale=small size, Pre-training Corpus=LS-960, Upstream=Proposed, #Upstream Params=32.54M, Downstream=Linear, #Downstream Params=0.13M, Evaluation Protocol=five-fold leave-one-session-out2023.1259.79
vq-wav2vecModel Scale=small size, Pre-training Corpus=LS-960, Upstream=Proposed, #Upstream Params=34.15M, Downstream=Linear, #Downstream Params=0.20M, Evaluation Protocol=five-fold leave-one-session-out2023.1258.24
SSAST#Param. (M)=89, Data (hrs)=5,800, Evaluation Protocol=Fine-tuning2026.0455.21