Experiment 使用不同language model的差別
%WER 50.81 [ 15640 / 30780, 1584 ins, 6817 del, 7239 sub ] exp_DAAIKidFullDeceptCSRCformosa_all_Trans_ADOS_train_happynvalid/chain/ADOS_tdnn_fold_transfer_lrFac0.75_NEpoch1/decode_ADOS_afraidangrysad/cer_8_0.0
%WER 54.93 [ 16908 / 30780, 1838 ins, 6394 del, 8676 sub ] exp_DAAIKidFullDeceptCSRCformosa_all_Trans_ADOS_train_happynvalid/chain/ADOS_tdnn_fold_transfer_lrFac0.75_NEpoch1/decode_ADOS_tdnn_fold_transfer_lrFac0.75_NEpoch1/cer_7_1.0 (使用只有ADOS_train_happynvalid的LM)
%WER 51.31 [ 15792 / 30780, 1937 ins, 5899 del, 7956 sub ] exp_DAAIKidFullDeceptCSRCformosa_all_Trans_ADOS_train_happynvalid/chain/ADOS_tdnn_fold_transfer_lrFac0.75_NEpoch1/decode_ADOS_tdnn_fold_transfer_lrFac0.75_NEpoch1/cer_8_0.0 (使用mix ADOS_train_happynvalid 和 formosa all 然後用ADOS_afraidangrysad 當測試集生出來的language model)
%WER 37.75 [ 11619 / 30780, 1268 ins, 6057 del, 4294 sub ] exp_DAAIKidFullDeceptCSRCformosa_all_Trans_ADOS_train_happynvalid/chain/ADOS_tdnn_fold_transfer_lrFac0.75_NEpoch1/decode_ADOS_tdnn_fold_transfer_lrFac0.75_NEpoch1/cer_9_0.0 (language model直接用ADOS_afraidangrysad訓練的language model)
Edited by JackingChen
