Speech Recognition on Hub5'00 (SWB, CHM)
4.3SWB ScoreSystem Combination (6+7+8+9) + TrafoXL LM
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| System Combination (6+7+8+9) + TrafoXL LMTraining Data=SWB-2k, External LM=LSTM+TrafoXL2021.05 | 4.3 | 6.8 | |
| Conformer Encoder-Decoder (id 9) + TrafoXL LMTraining Data=SWB-2k, Params=154M, Step Size=10ms, i-vector=false, External LM=LSTM+TrafoXL2021.05 | 4.5 | 7.2 | |
| LSTM Encoder (id 6)Training Data=SWB-2k, Params=661M, Step Size=10ms, i-vector=false, External LM=None2021.05 | 4.6 | 7.8 | |
| Conformer Encoder (id 8) + TrafoXL LMTraining Data=SWB-2k, Params=99M, Step Size=10ms, i-vector=false, External LM=LSTM+TrafoXL2021.05 | 4.6 | 7.6 | |
| Conformer Encoder (id 8)Training Data=SWB-2k, Params=99M, Step Size=10ms, i-vector=false, External LM=None2021.05 | 4.8 | 8 | |
| LSTM Encoder (id 7)Training Data=SWB-2k, Params=663M, Step Size=2.5ms, i-vector=false, External LM=None2021.05 | 4.9 | 7.7 | |
| System Combination (2+4+5) + TrafoXL LMTraining Data=SWB-300, External LM=LSTM+TrafoXL2021.05 | 5 | 10 | |
| System Combination (1+3+5) + TrafoXL LMTraining Data=SWB-300, External LM=LSTM+TrafoXL2021.05 | 5.1 | 10.1 | |
| Conformer Encoder (id 5) + TrafoXL LMTraining Data=SWB-300, Params=68M, Step Size=10ms, i-vector=false, External LM=LSTM+TrafoXL2021.05 | 5.5 | 11.2 | |
| LSTM Encoder-Decoder (id 4)Training Data=SWB-300, Params=60M, Step Size=2.5ms, i-vector=true, External LM=LSTM+TrafoXL2021.05 | 5.7 | 11.3 | |
| Conformer Encoder (id 5) + LSTM LMTraining Data=SWB-300, Params=68M, Step Size=10ms, i-vector=false, External LM=LSTM2021.05 | 5.7 | 11.4 | |
| LSTM Encoder (id 1) + LMTraining Data=SWB-300, Params=57M, Step Size=10ms, i-vector=false, External LM=LSTM2021.05 | 5.8 | 12.1 | |
| LSTM Encoder (id 3)Training Data=SWB-300, Params=56M, Step Size=2.5ms, i-vector=true, External LM=None2021.05 | 5.9 | 11.5 | |
| LSTM Encoder (id 1)Training Data=SWB-300, Params=57M, Step Size=10ms, i-vector=false, External LM=None2021.05 | 6 | 12 | |
| LSTM Encoder (id 2)Training Data=SWB-300, Params=61M, Step Size=10ms, i-vector=true, External LM=None2021.05 | 6 | 12 | |
| System Combination (1+3+5)Training Data=SWB-300, External LM=None2021.05 | 6.1 | 11.9 | |
| Conformer Encoder (id 5)Training Data=SWB-300, Params=68M, Step Size=10ms, i-vector=false, External LM=None2021.05 | 6.7 | 13 | |
| LF-MMIEnd-to-end=false, Language Model=RNN2019.04 | 7.3 | 14.2 | |
| CTC + Gram-CTCEnd-to-end=true, Language Model=N-gram2019.04 | 7.3 | 14.7 | |
| Phone E2E LF-MMIEnd-to-end=true, Language Model=RNN2019.04 | 7.5 | 14.6 | |
| Jasper DR 10x5End-to-end=true, Language Model=Transformer-XL, optimizer=SGD with momentum2019.04 | 7.8 | 16.2 | |
| Char E2E LF-MMIEnd-to-end=true, Language Model=RNN2019.04 | 8 | 17.6 | |
| RNN-TEnd-to-end=true, Language Model=4-gram2019.04 | 8.1 | 17.5 | |
| Attention Seq2SeqEnd-to-end=true2019.04 | 8.3 | 15.5 | |
| Jasper DR 10x5End-to-end=true, Language Model=4-gram, optimizer=SGD with momentum2019.04 | 8.3 | 19.3 |