GitLab

實驗： 轉移後注音系統 （langMapped），data為Formosa(all) DAAIKid(all) Deceipt

---

language model: lang_test_mapped

沒調參數

%WER 79.20 [ 24340 / 30731, 1149 ins, 9696 del, 13495 sub ] exp_langMapped_PureFormosaallDAAIKidallDeceipt/tri1/decode_ADOS_afraidangrysad/cer_11_0.5
%WER 77.17 [ 23714 / 30731, 1157 ins, 9536 del, 13021 sub ] exp_langMapped_PureFormosaallDAAIKidallDeceipt/tri2/decode_ADOS_afraidangrysad/cer_11_0.0
%WER 77.65 [ 23864 / 30731, 1295 ins, 10052 del, 12517 sub ] exp_langMapped_PureFormosaallDAAIKidallDeceipt/tri3a/decode_ADOS_afraidangrysad/cer_11_0.5
%WER **73.03** [ 22442 / 30731, 1572 ins, 8291 del, 12579 sub ] exp_langMapped_PureFormosaallDAAIKidallDeceipt/tri4a/decode_ADOS_afraidangrysad/cer_10_0.0
%WER 77.78 [ 23904 / 30731, 1348 ins, 9139 del, 13417 sub ] exp_langMapped_PureFormosaallDAAIKidallDeceipt/tri4a/decode_ADOS_afraidangrysad.si/cer_8_0.0
%WER 73.49 [ 22583 / 30731, 1372 ins, 9081 del, 12130 sub ] exp_langMapped_PureFormosaallDAAIKidallDeceipt/tri5a/decode_ADOS_afraidangrysad/cer_12_0.0
%WER 77.85 [ 23925 / 30731, 1259 ins, 9564 del, 13102 sub ] exp_langMapped_PureFormosaallDAAIKidallDeceipt/tri5a/decode_ADOS_afraidangrysad.si/cer_9_0.0

調過參數, 使用資料： 從頭~tri2 用formosa, tri3開始～ tri5a用ADOS_train_happynvalid_formosa，tri6開始用 data/ADOS_train_happynvalid_formosa_DAAI_Deceipt_CSRCKID

language model: lang_test_mapped

%WER 85.39 [ 26242 / 30731, 563 ins, 12812 del, 12867 sub ] exp_langMapped/mono_40_10000/decode_ADOS_afraidangrysad/cer_9_0.0
%WER 80.91 [ 24863 / 30731, 995 ins, 11825 del, 12043 sub ] exp_langMapped/tri1_iters40_gauss9000_numleaves3000/decode_ADOS_afraidangrysad/cer_11_0.0
%WER 79.49 [ 24429 / 30731, 997 ins, 11549 del, 11883 sub ] exp_langMapped/tri2_iters40_gauss16000_numleaves3000/decode_ADOS_afraidangrysad/cer_11_0.0
%WER 79.99 [ 24582 / 30731, 1297 ins, 10190 del, 13095 sub ] exp_langMapped/tri3a_iters40_gauss38000_numleaves9000_contexts3/decode_ADOS_afraidangrysad/cer_10_0.5
%WER 74.25 [ 22818 / 30731, 1483 ins, 8628 del, 12707 sub ] exp_langMapped/tri4a_gauss18000_numleaves5000/decode_ADOS_afraidangrysad/cer_13_0.0
%WER 80.41 [ 24712 / 30731, 1330 ins, 9644 del, 13738 sub ] exp_langMapped/tri4a_gauss18000_numleaves5000/decode_ADOS_afraidangrysad.si/cer_10_0.5
%WER 72.94 [ 22415 / 30731, 1584 ins, 8002 del, 12829 sub ] exp_langMapped/tri5a_gauss59000_numleaves3000/decode_ADOS_afraidangrysad/cer_10_1.0
%WER 78.53 [ 24133 / 30731, 1573 ins, 8677 del, 13883 sub ] exp_langMapped/tri5a_gauss59000_numleaves3000/decode_ADOS_afraidangrysad.si/cer_9_0.5
%WER **71.06** [ 21838 / 30731, 1725 ins, 6909 del, 13204 sub ] exp_langMapped/tri6a_gauss77000_numleaves3000_PureFormosaallDAAIKidallDeceipt/decode_ADOS_afraidangrysad/cer_10_0.5
%WER 77.74 [ 23891 / 30731, 1463 ins, 8265 del, 14163 sub ] exp_langMapped/tri6a_gauss77000_numleaves3000_PureFormosaallDAAIKidallDeceipt/decode_ADOS_afraidangrysad.si/cer_8_1.0
%WER 47.10 [ 14498 / 30780, 1667 ins, 5979 del, 6852 sub ] exp_langMapped_PureFormosaallDAAIKidallDeceipt/chain/model_20210127_ADOSFullDeceptDAAIKidFullformosaCSRC/decode_ADOS_afraidangrysad/cer_7_0.5

Chain model 的訓練圖(可以看出一個epoch就到平台了)：

比較BASELINE

---

原先沒轉換注音系統的狀況： langOrigin ，資料用Formosaall, DAAIKidall, Deceipt exp_langOrigin_PureFormosaallDAAIKidallDeceipt language model: lang_formosa

%WER 95.61 [ 29381 / 30731, 472 ins, 17214 del, 11695 sub ] exp_langOrigin_PureFormosaallDAAIKidallDeceipt/mono/decode_ADOS_afraidangrysad/cer_8_0.0
%WER 72.76 [ 22361 / 30731, 1163 ins, 9777 del, 11421 sub ] exp_langOrigin_PureFormosaallDAAIKidallDeceipt/tri1/decode_ADOS_afraidangrysad/cer_13_0.0
%WER 70.25 [ 21589 / 30731, 1060 ins, 9859 del, 10670 sub ] exp_langOrigin_PureFormosaallDAAIKidallDeceipt/tri2/decode_ADOS_afraidangrysad/cer_13_0.0
%WER 71.82 [ 22072 / 30731, 1133 ins, 10151 del, 10788 sub ] exp_langOrigin_PureFormosaallDAAIKidallDeceipt/tri3a/decode_ADOS_afraidangrysad/cer_13_0.0
%WER **65.95** [ 20266 / 30731, 1339 ins, 8830 del, 10097 sub ] exp_langOrigin_PureFormosaallDAAIKidallDeceipt/tri4a/decode_ADOS_afraidangrysad/cer_13_0.0
%WER 72.30 [ 22217 / 30731, 1199 ins, 9560 del, 11458 sub ] exp_langOrigin_PureFormosaallDAAIKidallDeceipt/tri4a/decode_ADOS_afraidangrysad.si/cer_10_0.0
%WER 67.93 [ 20876 / 30731, 1926 ins, 8192 del, 10758 sub ] exp_langOrigin_PureFormosaallDAAIKidallDeceipt/tri5a/decode_ADOS_afraidangrysad/cer_13_0.0
%WER 73.01 [ 22437 / 30731, 1182 ins, 9848 del, 11407 sub ] exp_langOrigin_PureFormosaallDAAIKidallDeceipt/tri5a/decode_ADOS_afraidangrysad.si/cer_10_0.5

# 這是拿以前的
%WER 37.14 [ 11412 / 30731, 1751 ins, 4480 del, 5181 sub ] KaldiBestModel20201230/exp_ADOShappyDAAIKidallDeceiptformosaCSRC/chain/ADOS_tdnn_fold_transfer/decode_ADOS_afraidangrysad/cer_9_0.0

Chain model 的訓練圖(可以看出一個epoch就到平台了)：