List your company

Phoneme Recognition on TIMIT core (test)

10.3PERTrain from oracle
9.20816.57923.9531.321Mar 22, 2013Jan 13, 2015Nov 5, 2016Aug 29, 2018Jun 21, 2020Apr 14, 2022Feb 6, 2024

Evaluation results

MethodLinks
Train from oracleGreedy decoding=true2024.0210.3—
REBORNGreedy decoding=true2024.0213.5—
EUROGreedy decoding=false, Language Model=4-gram, Backbone=WavLM2024.0214.6—
HMM over Time and Frequency Convolutional Net2015.0616.7—
Baseline + Conv. Features + Smooth FocusConvolutional Attention Features=true, Smooth Focus=true2015.0617.6—
PRETRANS-3L-250HWeights=4.3M, Training Method=Pretrained RNN Transducer, Number of Layers=3, Hidden Units=250, Cell Type=LSTM, Bidirectional=true2013.0317.7144
RNN Transducer2015.0617.7—
wav2vec-U+WFSTGreedy decoding=false, Language Model=4-gram2024.0217.8—
Baseline + Conv. FeaturesConvolutional Attention Features=true2015.0618—
TRANS-3L-250HWeights=4.3M, Training Method=RNN Transducer, Number of Layers=3, Hidden Units=250, Cell Type=LSTM, Bidirectional=true2013.0318.3112
CTC-5L-250HWeights=6.8M, Cell Type=LSTM, Number of Layers=5, Hidden Units=250, Bidirectional=true2013.0318.4150
CTC-3L-250HWeights=3.8M, Cell Type=LSTM, Number of Layers=3, Hidden Units=250, Bidirectional=true2013.0318.6124
Baseline Model2015.0618.7—
CTC-3L-421H-UNIWeights=3.8M, Cell Type=LSTM, Number of Layers=3, Hidden Units=421, Bidirectional=false2013.0319.6115
EUROGreedy decoding=false, Language Model=4-gram, Backbone=wav2vec 2.02024.0219.8—
CTC-2L-250HWeights=2.3M, Cell Type=LSTM, Number of Layers=2, Hidden Units=250, Bidirectional=true2013.032155
wav2vec-U (reproduced)Greedy decoding=true2024.0222.2—
CTC-1L-622HWeights=3.8M, Cell Type=LSTM, Number of Layers=1, Hidden Units=622, Bidirectional=true2013.032387
CTC-1L-250HWeights=0.8M, Cell Type=LSTM, Number of Layers=1, Hidden Units=250, Bidirectional=true2013.0323.982
CTC-3L-500H-TANHWeights=3.7M, Cell Type=tanh, Number of Layers=3, Hidden Units=500, Bidirectional=true2013.0337.6107