Blame view
egs/csj/s5/RESULTS
20 KB
8dcb6dfcb
first commit
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 |
#!/bin/bash
# This RESULTS file was obtained in Dec 4th 2017 in JHU. To see results from
# the scripts local/nnet3/ and local/chain/, look at the top of those files.
# Here, we just update the results of "chain nnet3" models under different data
# prepare modes. So the results of "nnet1" and "nnet3 tdnn" are a little bit
# out-of-date.
##### These experiments use default <number_mode> in run.sh "local/csj_data_prep.sh data/csj-data"
#i.e. Using "Academic lecture" and "other" data
for eval_num in `seq 3`; do echo "=== evaluation set $eval_num ===" ; for x in exp/{tri,dnn}*/decode_eval${eval_num}*; do [ -d $x ] && grep WER $x/wer_* | utils/best_wer.sh; done ; for x in exp/chain*/tdnn*/decode_eval${eval_num}*; do [ -d $x ] && grep WER $x/wer_* | utils/best_wer.sh; done ; done > RESULTS.ricky
exit 0
=== evaluation set 1 ===
%WER 22.04 [ 5736 / 26028, 473 ins, 1471 del, 3792 sub ] exp/tri1/decode_eval1_csj/wer_12_0.0
%WER 21.05 [ 5479 / 26028, 439 ins, 1459 del, 3581 sub ] exp/tri2/decode_eval1_csj/wer_14_0.0
%WER 16.96 [ 4414 / 26028, 472 ins, 1008 del, 2934 sub ] exp/tri3/decode_eval1_csj/wer_16_0.0
%WER 18.30 [ 4763 / 26028, 411 ins, 1205 del, 3147 sub ] exp/tri3/decode_eval1_csj_nosp/wer_16_0.0
%WER 14.90 [ 3879 / 26028, 457 ins, 876 del, 2546 sub ] exp/tri4/decode_eval1_csj/wer_17_0.0
%WER 17.14 [ 4460 / 26028, 516 ins, 971 del, 2973 sub ] exp/tri4/decode_eval1_csj.si/wer_17_0.0
%WER 14.09 [ 3668 / 26028, 429 ins, 807 del, 2432 sub ] exp/tri4_fmmi_b0.1/decode_eval1_it4_csj/wer_16_0.0
%WER 13.67 [ 3559 / 26028, 395 ins, 799 del, 2365 sub ] exp/tri4_fmmi_b0.1/decode_eval1_it5_csj/wer_17_0.0
%WER 13.39 [ 3486 / 26028, 353 ins, 831 del, 2302 sub ] exp/tri4_fmmi_b0.1/decode_eval1_it6_csj/wer_14_0.5
%WER 13.35 [ 3476 / 26028, 343 ins, 848 del, 2285 sub ] exp/tri4_fmmi_b0.1/decode_eval1_it7_csj/wer_15_0.5
%WER 13.31 [ 3465 / 26028, 368 ins, 815 del, 2282 sub ] exp/tri4_fmmi_b0.1/decode_eval1_it8_csj/wer_14_0.5
%WER 14.04 [ 3654 / 26028, 417 ins, 819 del, 2418 sub ] exp/tri4_mmi_b0.1/decode_eval1_1.mdl_csj/wer_16_0.0
%WER 13.77 [ 3585 / 26028, 408 ins, 801 del, 2376 sub ] exp/tri4_mmi_b0.1/decode_eval1_2.mdl_csj/wer_16_0.0
%WER 13.64 [ 3550 / 26028, 356 ins, 858 del, 2336 sub ] exp/tri4_mmi_b0.1/decode_eval1_3.mdl_csj/wer_14_0.5
%WER 13.52 [ 3519 / 26028, 342 ins, 878 del, 2299 sub ] exp/tri4_mmi_b0.1/decode_eval1_4.mdl_csj/wer_15_0.5
%WER 10.30 [ 2680 / 26028, 288 ins, 538 del, 1854 sub ] exp/chain/tdnn1a/decode_eval1/wer_12_0.0
%WER 10.30 [ 2681 / 26028, 292 ins, 538 del, 1851 sub ] exp/chain/tdnn1a_online/decode_eval1/wer_12_0.0
=== evaluation set 2 ===
%WER 18.68 [ 4980 / 26661, 540 ins, 998 del, 3442 sub ] exp/tri1/decode_eval2_csj/wer_13_0.0
%WER 17.62 [ 4697 / 26661, 555 ins, 837 del, 3305 sub ] exp/tri2/decode_eval2_csj/wer_12_0.0
%WER 14.46 [ 3856 / 26661, 535 ins, 630 del, 2691 sub ] exp/tri3/decode_eval2_csj/wer_15_0.0
%WER 15.31 [ 4082 / 26661, 498 ins, 743 del, 2841 sub ] exp/tri3/decode_eval2_csj_nosp/wer_15_0.0
%WER 12.43 [ 3315 / 26661, 410 ins, 592 del, 2313 sub ] exp/tri4/decode_eval2_csj/wer_16_0.5
%WER 17.38 [ 4633 / 26661, 716 ins, 650 del, 3267 sub ] exp/tri4/decode_eval2_csj.si/wer_17_0.0
%WER 11.41 [ 3041 / 26661, 324 ins, 595 del, 2122 sub ] exp/tri4_fmmi_b0.1/decode_eval2_it4_csj/wer_17_0.5
%WER 11.14 [ 2969 / 26661, 332 ins, 538 del, 2099 sub ] exp/tri4_fmmi_b0.1/decode_eval2_it5_csj/wer_15_0.5
%WER 11.09 [ 2958 / 26661, 322 ins, 545 del, 2091 sub ] exp/tri4_fmmi_b0.1/decode_eval2_it6_csj/wer_12_1.0
%WER 11.06 [ 2949 / 26661, 363 ins, 512 del, 2074 sub ] exp/tri4_fmmi_b0.1/decode_eval2_it7_csj/wer_14_0.5
%WER 11.02 [ 2937 / 26661, 286 ins, 580 del, 2071 sub ] exp/tri4_fmmi_b0.1/decode_eval2_it8_csj/wer_14_1.0
%WER 11.32 [ 3018 / 26661, 313 ins, 592 del, 2113 sub ] exp/tri4_mmi_b0.1/decode_eval2_1.mdl_csj/wer_17_0.5
%WER 11.13 [ 2968 / 26661, 329 ins, 542 del, 2097 sub ] exp/tri4_mmi_b0.1/decode_eval2_2.mdl_csj/wer_15_0.5
%WER 11.04 [ 2944 / 26661, 279 ins, 592 del, 2073 sub ] exp/tri4_mmi_b0.1/decode_eval2_3.mdl_csj/wer_14_1.0
%WER 11.08 [ 2955 / 26661, 301 ins, 567 del, 2087 sub ] exp/tri4_mmi_b0.1/decode_eval2_4.mdl_csj/wer_13_1.0
%WER 8.59 [ 2291 / 26661, 190 ins, 442 del, 1659 sub ] exp/chain/tdnn1a/decode_eval2/wer_13_0.5
%WER 8.56 [ 2282 / 26661, 213 ins, 375 del, 1694 sub ] exp/chain/tdnn1a_online/decode_eval2/wer_11_0.5
=== evaluation set 3 ===
%WER 23.73 [ 4079 / 17189, 465 ins, 867 del, 2747 sub ] exp/tri1/decode_eval3_csj/wer_15_0.0
%WER 23.08 [ 3968 / 17189, 518 ins, 772 del, 2678 sub ] exp/tri2/decode_eval3_csj/wer_14_0.0
%WER 18.59 [ 3196 / 17189, 403 ins, 639 del, 2154 sub ] exp/tri3/decode_eval3_csj/wer_17_0.5
%WER 20.00 [ 3438 / 17189, 488 ins, 642 del, 2308 sub ] exp/tri3/decode_eval3_csj_nosp/wer_16_0.0
%WER 15.99 [ 2749 / 17189, 342 ins, 550 del, 1857 sub ] exp/tri4/decode_eval3_csj/wer_17_1.0
%WER 20.45 [ 3515 / 17189, 448 ins, 671 del, 2396 sub ] exp/tri4/decode_eval3_csj.si/wer_17_1.0
%WER 15.06 [ 2589 / 17189, 307 ins, 516 del, 1766 sub ] exp/tri4_fmmi_b0.1/decode_eval3_it4_csj/wer_16_1.0
%WER 14.75 [ 2536 / 17189, 341 ins, 478 del, 1717 sub ] exp/tri4_fmmi_b0.1/decode_eval3_it5_csj/wer_17_0.5
%WER 14.54 [ 2499 / 17189, 300 ins, 502 del, 1697 sub ] exp/tri4_fmmi_b0.1/decode_eval3_it6_csj/wer_15_1.0
%WER 14.58 [ 2507 / 17189, 310 ins, 499 del, 1698 sub ] exp/tri4_fmmi_b0.1/decode_eval3_it7_csj/wer_15_1.0
%WER 14.58 [ 2507 / 17189, 333 ins, 479 del, 1695 sub ] exp/tri4_fmmi_b0.1/decode_eval3_it8_csj/wer_14_1.0
%WER 15.09 [ 2593 / 17189, 301 ins, 532 del, 1760 sub ] exp/tri4_mmi_b0.1/decode_eval3_1.mdl_csj/wer_16_1.0
%WER 14.84 [ 2551 / 17189, 352 ins, 475 del, 1724 sub ] exp/tri4_mmi_b0.1/decode_eval3_2.mdl_csj/wer_17_0.5
%WER 14.75 [ 2535 / 17189, 296 ins, 522 del, 1717 sub ] exp/tri4_mmi_b0.1/decode_eval3_3.mdl_csj/wer_16_1.0
%WER 14.93 [ 2566 / 17189, 306 ins, 550 del, 1710 sub ] exp/tri4_mmi_b0.1/decode_eval3_4.mdl_csj/wer_17_1.0
%WER 9.90 [ 1701 / 17189, 241 ins, 265 del, 1195 sub ] exp/chain/tdnn1a/decode_eval3/wer_13_0.0
%WER 9.90 [ 1702 / 17189, 242 ins, 265 del, 1195 sub ] exp/chain/tdnn1a_online/decode_eval3/wer_13_0.0
# nnet1 DNN
=== evaluation set 1 ===
%WER 11.69 [ 3042 / 26028, 325 ins, 711 del, 2006 sub ] exp/dnn5b_pretrain-dbn_dnn/decode_eval1_csj/wer_13_0.0
%WER 11.10 [ 2890 / 26028, 300 ins, 678 del, 1912 sub ] exp/dnn5b_pretrain-dbn_dnn_smbr/decode_eval1_csj/wer_15_0.5
%WER 10.96 [ 2853 / 26028, 319 ins, 642 del, 1892 sub ] exp/dnn5b_pretrain-dbn_dnn_smbr_i1lats/decode_eval1_csj/wer_15_0.5
=== evaluation set 2 ===
%WER 8.95 [ 2387 / 26661, 249 ins, 419 del, 1719 sub ] exp/dnn5b_pretrain-dbn_dnn/decode_eval2_csj/wer_13_0.0
%WER 8.83 [ 2353 / 26661, 296 ins, 376 del, 1681 sub ] exp/dnn5b_pretrain-dbn_dnn_smbr/decode_eval2_csj/wer_11_1.0
%WER 8.78 [ 2340 / 26661, 262 ins, 407 del, 1671 sub ] exp/dnn5b_pretrain-dbn_dnn_smbr_i1lats/decode_eval2_csj/wer_13_1.0
=== evaluation set 3 ===
%WER 12.47 [ 2144 / 17189, 311 ins, 355 del, 1478 sub ] exp/dnn5b_pretrain-dbn_dnn/decode_eval3_csj/wer_15_0.0
%WER 11.97 [ 2058 / 17189, 281 ins, 341 del, 1436 sub ] exp/dnn5b_pretrain-dbn_dnn_smbr/decode_eval3_csj/wer_14_1.0
%WER 11.65 [ 2002 / 17189, 284 ins, 318 del, 1400 sub ] exp/dnn5b_pretrain-dbn_dnn_smbr_i1lats/decode_eval3_csj/wer_14_1.0
# nnet3 TDNN (former tdnn results)
%WER 11.34 [ 2952 / 26028, 286 ins, 750 del, 1916 sub ] exp/nnet3/nnet_tdnn/decode_eval1_csj/wer_12_0.5
%WER 9.07 [ 2417 / 26661, 276 ins, 414 del, 1727 sub ] exp/nnet3/nnet_tdnn/decode_eval2_csj/wer_12_0.0
%WER 11.46 [ 1970 / 17189, 233 ins, 400 del, 1337 sub ] exp/nnet3/nnet_tdnn/decode_eval3_csj/wer_14_0.5
##### These experiments use <number_mode> 2 in run.sh "local/csj_data_prep.sh data/csj-data 2"
#i.e. Using All data except for "dialog" data
for eval_num in `seq 3`; do echo "=== evaluation set $eval_num ===" ; for x in exp/{tri,dnn}*/decode_eval${eval_num}*; do [ -d $x ] && grep WER $x/wer_* | utils/best_wer.sh; done ; for x in exp/chain*/tdnn*/decode_eval${eval_num}*; do [ -d $x ] && grep WER $x/wer_* | utils/best_wer.sh; done ; done > RESULTS.ricky
=== evaluation set 1 ===
%WER 22.44 [ 5840 / 26028, 416 ins, 1598 del, 3826 sub ] exp/tri1/decode_eval1_csj/wer_14_0.0
%WER 21.22 [ 5523 / 26028, 430 ins, 1456 del, 3637 sub ] exp/tri2/decode_eval1_csj/wer_14_0.0
%WER 17.37 [ 4521 / 26028, 450 ins, 1079 del, 2992 sub ] exp/tri3/decode_eval1_csj/wer_16_0.0
%WER 18.83 [ 4901 / 26028, 378 ins, 1316 del, 3207 sub ] exp/tri3/decode_eval1_csj_nosp/wer_17_0.0
%WER 15.15 [ 3942 / 26028, 420 ins, 929 del, 2593 sub ] exp/tri4/decode_eval1_csj/wer_17_0.0
%WER 18.76 [ 4883 / 26028, 610 ins, 1016 del, 3257 sub ] exp/tri4/decode_eval1_csj.si/wer_16_0.0
%WER 14.38 [ 3742 / 26028, 319 ins, 978 del, 2445 sub ] exp/tri4_fmmi_b0.1/decode_eval1_it4_csj/wer_15_0.5
%WER 13.79 [ 3589 / 26028, 361 ins, 856 del, 2372 sub ] exp/tri4_fmmi_b0.1/decode_eval1_it5_csj/wer_16_0.0
%WER 13.65 [ 3553 / 26028, 329 ins, 902 del, 2322 sub ] exp/tri4_fmmi_b0.1/decode_eval1_it6_csj/wer_14_0.5
%WER 13.62 [ 3546 / 26028, 363 ins, 843 del, 2340 sub ] exp/tri4_fmmi_b0.1/decode_eval1_it7_csj/wer_13_0.5
%WER 13.66 [ 3556 / 26028, 408 ins, 801 del, 2347 sub ] exp/tri4_fmmi_b0.1/decode_eval1_it8_csj/wer_12_0.5
%WER 14.32 [ 3727 / 26028, 364 ins, 884 del, 2479 sub ] exp/tri4_mmi_b0.1/decode_eval1_1.mdl_csj/wer_16_0.0
%WER 14.15 [ 3683 / 26028, 383 ins, 869 del, 2431 sub ] exp/tri4_mmi_b0.1/decode_eval1_2.mdl_csj/wer_16_0.0
%WER 13.94 [ 3629 / 26028, 394 ins, 841 del, 2394 sub ] exp/tri4_mmi_b0.1/decode_eval1_3.mdl_csj/wer_16_0.0
%WER 13.89 [ 3614 / 26028, 432 ins, 787 del, 2395 sub ] exp/tri4_mmi_b0.1/decode_eval1_4.mdl_csj/wer_14_0.0
%WER 10.07 [ 2620 / 26028, 274 ins, 497 del, 1849 sub ] exp/chain/tdnn1a/decode_eval1/wer_12_0.0
%WER 10.01 [ 2606 / 26028, 255 ins, 524 del, 1827 sub ] exp/chain/tdnn1a_online/decode_eval1/wer_13_0.0
=== evaluation set 2 ===
%WER 18.49 [ 4930 / 26661, 507 ins, 1001 del, 3422 sub ] exp/tri1/decode_eval2_csj/wer_13_0.0
%WER 17.79 [ 4744 / 26661, 586 ins, 821 del, 3337 sub ] exp/tri2/decode_eval2_csj/wer_11_0.0
%WER 14.29 [ 3810 / 26661, 488 ins, 629 del, 2693 sub ] exp/tri3/decode_eval2_csj/wer_14_0.0
%WER 15.28 [ 4075 / 26661, 474 ins, 726 del, 2875 sub ] exp/tri3/decode_eval2_csj_nosp/wer_13_0.0
%WER 12.09 [ 3222 / 26661, 418 ins, 559 del, 2245 sub ] exp/tri4/decode_eval2_csj/wer_13_0.5
%WER 17.93 [ 4781 / 26661, 662 ins, 724 del, 3395 sub ] exp/tri4/decode_eval2_csj.si/wer_14_0.5
%WER 11.35 [ 3027 / 26661, 338 ins, 546 del, 2143 sub ] exp/tri4_fmmi_b0.1/decode_eval2_it4_csj/wer_13_0.5
%WER 11.04 [ 2944 / 26661, 332 ins, 540 del, 2072 sub ] exp/tri4_fmmi_b0.1/decode_eval2_it5_csj/wer_13_0.5
%WER 10.86 [ 2896 / 26661, 302 ins, 531 del, 2063 sub ] exp/tri4_fmmi_b0.1/decode_eval2_it6_csj/wer_11_1.0
%WER 10.86 [ 2895 / 26661, 313 ins, 535 del, 2047 sub ] exp/tri4_fmmi_b0.1/decode_eval2_it7_csj/wer_11_1.0
%WER 10.91 [ 2910 / 26661, 294 ins, 545 del, 2071 sub ] exp/tri4_fmmi_b0.1/decode_eval2_it8_csj/wer_12_1.0
%WER 11.19 [ 2983 / 26661, 380 ins, 489 del, 2114 sub ] exp/tri4_mmi_b0.1/decode_eval2_1.mdl_csj/wer_14_0.0
%WER 11.11 [ 2963 / 26661, 327 ins, 542 del, 2094 sub ] exp/tri4_mmi_b0.1/decode_eval2_2.mdl_csj/wer_13_0.5
%WER 10.97 [ 2924 / 26661, 297 ins, 542 del, 2085 sub ] exp/tri4_mmi_b0.1/decode_eval2_3.mdl_csj/wer_11_1.0
%WER 10.88 [ 2902 / 26661, 299 ins, 525 del, 2078 sub ] exp/tri4_mmi_b0.1/decode_eval2_4.mdl_csj/wer_11_1.0
%WER 8.38 [ 2234 / 26661, 200 ins, 382 del, 1652 sub ] exp/chain/tdnn1a/decode_eval2/wer_12_0.5
%WER 8.33 [ 2222 / 26661, 230 ins, 352 del, 1640 sub ] exp/chain/tdnn1a_online/decode_eval2/wer_13_0.0
=== evaluation set 3 ===
%WER 24.27 [ 4171 / 17189, 427 ins, 978 del, 2766 sub ] exp/tri1/decode_eval3_csj/wer_13_0.5
%WER 23.01 [ 3955 / 17189, 521 ins, 785 del, 2649 sub ] exp/tri2/decode_eval3_csj/wer_14_0.0
%WER 16.70 [ 2871 / 17189, 352 ins, 601 del, 1918 sub ] exp/tri3/decode_eval3_csj/wer_15_0.5
%WER 17.68 [ 3039 / 17189, 385 ins, 617 del, 2037 sub ] exp/tri3/decode_eval3_csj_nosp/wer_16_0.0
%WER 14.11 [ 2425 / 17189, 342 ins, 425 del, 1658 sub ] exp/tri4/decode_eval3_csj/wer_16_0.5
%WER 19.46 [ 3345 / 17189, 523 ins, 521 del, 2301 sub ] exp/tri4/decode_eval3_csj.si/wer_15_0.5
%WER 13.54 [ 2328 / 17189, 326 ins, 395 del, 1607 sub ] exp/tri4_fmmi_b0.1/decode_eval3_it4_csj/wer_15_0.5
%WER 12.94 [ 2224 / 17189, 291 ins, 420 del, 1513 sub ] exp/tri4_fmmi_b0.1/decode_eval3_it5_csj/wer_12_1.0
%WER 12.62 [ 2170 / 17189, 295 ins, 387 del, 1488 sub ] exp/tri4_fmmi_b0.1/decode_eval3_it6_csj/wer_11_1.0
%WER 12.61 [ 2167 / 17189, 300 ins, 388 del, 1479 sub ] exp/tri4_fmmi_b0.1/decode_eval3_it7_csj/wer_12_1.0
%WER 12.58 [ 2163 / 17189, 302 ins, 384 del, 1477 sub ] exp/tri4_fmmi_b0.1/decode_eval3_it8_csj/wer_12_1.0
%WER 13.13 [ 2257 / 17189, 315 ins, 406 del, 1536 sub ] exp/tri4_mmi_b0.1/decode_eval3_1.mdl_csj/wer_15_0.5
%WER 12.91 [ 2219 / 17189, 320 ins, 390 del, 1509 sub ] exp/tri4_mmi_b0.1/decode_eval3_2.mdl_csj/wer_14_0.5
%WER 12.80 [ 2201 / 17189, 288 ins, 406 del, 1507 sub ] exp/tri4_mmi_b0.1/decode_eval3_3.mdl_csj/wer_12_1.0
%WER 12.79 [ 2198 / 17189, 359 ins, 343 del, 1496 sub ] exp/tri4_mmi_b0.1/decode_eval3_4.mdl_csj/wer_12_0.5
%WER 7.94 [ 1365 / 17189, 133 ins, 251 del, 981 sub ] exp/chain/tdnn1a/decode_eval3/wer_11_1.0
%WER 7.89 [ 1356 / 17189, 173 ins, 211 del, 972 sub ] exp/chain/tdnn1a_online/decode_eval3/wer_11_0.5
# nnet3 TDNN (former tdnn results)
%WER 11.15 [ 2901 / 26028, 297 ins, 697 del, 1907 sub ] exp/nnet3/nnet_tdnn/decode_eval1_csj/wer_15_0.0
%WER 8.47 [ 2259 / 26661, 289 ins, 381 del, 1589 sub ] exp/nnet3/nnet_tdnn/decode_eval2_csj/wer_12_0.0
%WER 9.21 [ 1583 / 17189, 240 ins, 258 del, 1085 sub ] exp/nnet3/nnet_tdnn/decode_eval3_csj/wer_12_0.0
##### These experiments use <number_mode> 3 in run.sh "local/csj_data_prep.sh data/csj-data 3"
#i.e. Using All data
for eval_num in `seq 3`; do echo "=== evaluation set $eval_num ===" ; for x in exp/{tri,dnn}*/decode_eval${eval_num}*; do [ -d $x ] && grep WER $x/wer_* | utils/best_wer.sh; done ; for x in exp/chain*/tdnn*/decode_eval${eval_num}*; do [ -d $x ] && grep WER $x/wer_* | utils/best_wer.sh; done ; done > RESULTS.ricky_exp
=== evaluation set 1 ===
%WER 22.44 [ 5841 / 26028, 416 ins, 1598 del, 3827 sub ] exp/tri1/decode_eval1_csj/wer_14_0.0
%WER 21.19 [ 5516 / 26028, 454 ins, 1398 del, 3664 sub ] exp/tri2/decode_eval1_csj/wer_13_0.0
%WER 17.39 [ 4527 / 26028, 446 ins, 1106 del, 2975 sub ] exp/tri3/decode_eval1_csj/wer_17_0.0
%WER 18.73 [ 4874 / 26028, 414 ins, 1236 del, 3224 sub ] exp/tri3/decode_eval1_csj_nosp/wer_15_0.0
%WER 15.08 [ 3926 / 26028, 405 ins, 923 del, 2598 sub ] exp/tri4/decode_eval1_csj/wer_17_0.0
%WER 18.49 [ 4813 / 26028, 604 ins, 985 del, 3224 sub ] exp/tri4/decode_eval1_csj.si/wer_14_0.0
%WER 14.47 [ 3766 / 26028, 395 ins, 856 del, 2515 sub ] exp/tri4_fmmi_b0.1/decode_eval1_it4_csj/wer_15_0.0
%WER 14.01 [ 3646 / 26028, 397 ins, 806 del, 2443 sub ] exp/tri4_fmmi_b0.1/decode_eval1_it5_csj/wer_14_0.0
%WER 13.89 [ 3615 / 26028, 423 ins, 787 del, 2405 sub ] exp/tri4_fmmi_b0.1/decode_eval1_it6_csj/wer_14_0.0
%WER 13.61 [ 3543 / 26028, 329 ins, 876 del, 2338 sub ] exp/tri4_fmmi_b0.1/decode_eval1_it7_csj/wer_14_0.5
%WER 13.60 [ 3539 / 26028, 390 ins, 802 del, 2347 sub ] exp/tri4_fmmi_b0.1/decode_eval1_it8_csj/wer_12_0.5
%WER 14.29 [ 3720 / 26028, 379 ins, 878 del, 2463 sub ] exp/tri4_mmi_b0.1/decode_eval1_1.mdl_csj/wer_16_0.0
%WER 14.06 [ 3660 / 26028, 375 ins, 862 del, 2423 sub ] exp/tri4_mmi_b0.1/decode_eval1_2.mdl_csj/wer_16_0.0
%WER 14.03 [ 3651 / 26028, 421 ins, 800 del, 2430 sub ] exp/tri4_mmi_b0.1/decode_eval1_3.mdl_csj/wer_14_0.0
%WER 13.89 [ 3616 / 26028, 453 ins, 754 del, 2409 sub ] exp/tri4_mmi_b0.1/decode_eval1_4.mdl_csj/wer_13_0.0
%WER 10.10 [ 2630 / 26028, 261 ins, 515 del, 1854 sub ] exp/chain/tdnn1a/decode_eval1/wer_12_0.0
%WER 10.11 [ 2632 / 26028, 243 ins, 546 del, 1843 sub ] exp/chain/tdnn1a_online/decode_eval1/wer_13_0.0
=== evaluation set 2 ===
%WER 18.49 [ 4929 / 26661, 506 ins, 1001 del, 3422 sub ] exp/tri1/decode_eval2_csj/wer_13_0.0
%WER 17.46 [ 4656 / 26661, 522 ins, 894 del, 3240 sub ] exp/tri2/decode_eval2_csj/wer_13_0.0
%WER 14.45 [ 3853 / 26661, 508 ins, 641 del, 2704 sub ] exp/tri3/decode_eval2_csj/wer_14_0.0
%WER 15.20 [ 4053 / 26661, 441 ins, 793 del, 2819 sub ] exp/tri3/decode_eval2_csj_nosp/wer_15_0.0
%WER 12.22 [ 3259 / 26661, 365 ins, 655 del, 2239 sub ] exp/tri4/decode_eval2_csj/wer_17_0.5
%WER 17.49 [ 4663 / 26661, 610 ins, 742 del, 3311 sub ] exp/tri4/decode_eval2_csj.si/wer_14_0.5
%WER 11.45 [ 3054 / 26661, 422 ins, 503 del, 2129 sub ] exp/tri4_fmmi_b0.1/decode_eval2_it4_csj/wer_14_0.0
%WER 10.93 [ 2914 / 26661, 327 ins, 543 del, 2044 sub ] exp/tri4_fmmi_b0.1/decode_eval2_it5_csj/wer_13_0.5
%WER 10.72 [ 2858 / 26661, 270 ins, 584 del, 2004 sub ] exp/tri4_fmmi_b0.1/decode_eval2_it6_csj/wer_12_1.0
%WER 10.79 [ 2877 / 26661, 353 ins, 492 del, 2032 sub ] exp/tri4_fmmi_b0.1/decode_eval2_it7_csj/wer_12_0.5
%WER 10.75 [ 2867 / 26661, 282 ins, 570 del, 2015 sub ] exp/tri4_fmmi_b0.1/decode_eval2_it8_csj/wer_12_1.0
%WER 11.14 [ 2969 / 26661, 321 ins, 574 del, 2074 sub ] exp/tri4_mmi_b0.1/decode_eval2_1.mdl_csj/wer_14_0.5
%WER 10.92 [ 2912 / 26661, 338 ins, 523 del, 2051 sub ] exp/tri4_mmi_b0.1/decode_eval2_2.mdl_csj/wer_12_0.5
%WER 10.83 [ 2888 / 26661, 326 ins, 512 del, 2050 sub ] exp/tri4_mmi_b0.1/decode_eval2_3.mdl_csj/wer_12_0.5
%WER 10.79 [ 2878 / 26661, 333 ins, 497 del, 2048 sub ] exp/tri4_mmi_b0.1/decode_eval2_4.mdl_csj/wer_12_0.5
%WER 8.08 [ 2153 / 26661, 214 ins, 340 del, 1599 sub ] exp/chain/tdnn1a/decode_eval2/wer_11_0.5
%WER 8.06 [ 2149 / 26661, 216 ins, 334 del, 1599 sub ] exp/chain/tdnn1a_online/decode_eval2/wer_11_0.5
=== evaluation set 3 ===
%WER 24.27 [ 4171 / 17189, 427 ins, 976 del, 2768 sub ] exp/tri1/decode_eval3_csj/wer_13_0.5
%WER 22.90 [ 3937 / 17189, 535 ins, 726 del, 2676 sub ] exp/tri2/decode_eval3_csj/wer_13_0.0
%WER 16.52 [ 2840 / 17189, 325 ins, 608 del, 1907 sub ] exp/tri3/decode_eval3_csj/wer_16_0.5
%WER 17.42 [ 2994 / 17189, 410 ins, 571 del, 2013 sub ] exp/tri3/decode_eval3_csj_nosp/wer_14_0.0
%WER 14.20 [ 2441 / 17189, 389 ins, 388 del, 1664 sub ] exp/tri4/decode_eval3_csj/wer_17_0.0
%WER 19.01 [ 3268 / 17189, 514 ins, 496 del, 2258 sub ] exp/tri4/decode_eval3_csj.si/wer_14_0.5
%WER 13.59 [ 2336 / 17189, 335 ins, 398 del, 1603 sub ] exp/tri4_fmmi_b0.1/decode_eval3_it4_csj/wer_15_0.5
%WER 12.99 [ 2233 / 17189, 272 ins, 436 del, 1525 sub ] exp/tri4_fmmi_b0.1/decode_eval3_it5_csj/wer_13_1.0
%WER 12.80 [ 2201 / 17189, 361 ins, 323 del, 1517 sub ] exp/tri4_fmmi_b0.1/decode_eval3_it6_csj/wer_11_0.5
%WER 12.75 [ 2192 / 17189, 335 ins, 364 del, 1493 sub ] exp/tri4_fmmi_b0.1/decode_eval3_it7_csj/wer_13_0.5
%WER 12.79 [ 2198 / 17189, 321 ins, 389 del, 1488 sub ] exp/tri4_fmmi_b0.1/decode_eval3_it8_csj/wer_14_0.5
%WER 13.24 [ 2275 / 17189, 301 ins, 427 del, 1547 sub ] exp/tri4_mmi_b0.1/decode_eval3_1.mdl_csj/wer_16_0.5
%WER 12.97 [ 2230 / 17189, 308 ins, 419 del, 1503 sub ] exp/tri4_mmi_b0.1/decode_eval3_2.mdl_csj/wer_15_0.5
%WER 12.97 [ 2229 / 17189, 354 ins, 358 del, 1517 sub ] exp/tri4_mmi_b0.1/decode_eval3_3.mdl_csj/wer_12_0.5
%WER 12.97 [ 2230 / 17189, 348 ins, 360 del, 1522 sub ] exp/tri4_mmi_b0.1/decode_eval3_4.mdl_csj/wer_12_0.5
%WER 7.78 [ 1338 / 17189, 171 ins, 200 del, 967 sub ] exp/chain/tdnn1a/decode_eval3/wer_11_0.5
%WER 7.84 [ 1347 / 17189, 156 ins, 215 del, 976 sub ] exp/chain/tdnn1a_online/decode_eval3/wer_10_1.0
# nnet1 DNN
=== evaluation set 1 ===
%WER 11.64 [ 3030 / 26028, 325 ins, 703 del, 2002 sub ] exp/dnn5b_pretrain-dbn_dnn/decode_eval1_csj/wer_12_0.0
%WER 10.76 [ 2800 / 26028, 276 ins, 644 del, 1880 sub ] exp/dnn5b_pretrain-dbn_dnn_smbr/decode_eval1_csj/wer_12_1.0
%WER 10.66 [ 2774 / 26028, 290 ins, 643 del, 1841 sub ] exp/dnn5b_pretrain-dbn_dnn_smbr_i1lats/decode_eval1_csj/wer_13_1.0
=== evaluation set 2 ===
%WER 8.77 [ 2338 / 26661, 246 ins, 425 del, 1667 sub ] exp/dnn5b_pretrain-dbn_dnn/decode_eval2_csj/wer_12_0.0
%WER 8.60 [ 2293 / 26661, 239 ins, 447 del, 1607 sub ] exp/dnn5b_pretrain-dbn_dnn_smbr/decode_eval2_csj/wer_14_1.0
%WER 8.46 [ 2255 / 26661, 289 ins, 354 del, 1612 sub ] exp/dnn5b_pretrain-dbn_dnn_smbr_i1lats/decode_eval2_csj/wer_11_1.0
=== evaluation set 3 ===
%WER 9.92 [ 1705 / 17189, 254 ins, 264 del, 1187 sub ] exp/dnn5b_pretrain-dbn_dnn/decode_eval3_csj/wer_11_0.0
%WER 9.41 [ 1618 / 17189, 226 ins, 269 del, 1123 sub ] exp/dnn5b_pretrain-dbn_dnn_smbr/decode_eval3_csj/wer_14_1.0
%WER 9.12 [ 1568 / 17189, 264 ins, 217 del, 1087 sub ] exp/dnn5b_pretrain-dbn_dnn_smbr_i1lats/decode_eval3_csj/wer_13_0.5
|