[WIP] This branch contains some acoustic modeling improvements... still cleaning up.

[danpovey]

No description provided.

[davidavdav]

Hello,

I've run this script, or at least, the nnet configuration, in a Librispeech set-up, and results are:

exp/chain/tdnn7m23t_sp/decode_dev_clean_fglarge/wer_12_0.5:%WER 3.40 [ 1851 / 54402, 229 ins, 155 del, 1467 sub ]
exp/chain/tdnn7m23t_sp/decode_dev_clean_tglarge/wer_12_1.0:%WER 3.51 [ 1910 / 54402, 202 ins, 199 del, 1509 sub ]
exp/chain/tdnn7m23t_sp/decode_dev_clean_tgmed/wer_12_0.0:%WER 4.28 [ 2327 / 54402, 265 ins, 206 del, 1856 sub ]
exp/chain/tdnn7m23t_sp/decode_dev_clean_tgsmall/wer_12_0.0:%WER 4.84 [ 2635 / 54402, 277 ins, 263 del, 2095 sub ]
exp/chain/tdnn7m23t_sp/decode_dev_other_fglarge/wer_14_0.5:%WER 8.77 [ 4468 / 50948, 484 ins, 500 del, 3484 sub ]
exp/chain/tdnn7m23t_sp/decode_dev_other_tglarge/wer_14_0.0:%WER 9.24 [ 4709 / 50948, 598 ins, 435 del, 3676 sub ]
exp/chain/tdnn7m23t_sp/decode_dev_other_tgmed/wer_14_0.0:%WER 11.31 [ 5760 / 50948, 604 ins, 678 del, 4478 sub ]
exp/chain/tdnn7m23t_sp/decode_dev_other_tgsmall/wer_14_0.0:%WER 12.45 [ 6342 / 50948, 596 ins, 829 del, 4917 sub ]
exp/chain/tdnn7m23t_sp/decode_test_clean_fglarge/wer_11_0.5:%WER 3.87 [ 2036 / 52576, 317 ins, 157 del, 1562 sub ]
exp/chain/tdnn7m23t_sp/decode_test_clean_tglarge/wer_10_0.5:%WER 4.04 [ 2126 / 52576, 336 ins, 164 del, 1626 sub ]
exp/chain/tdnn7m23t_sp/decode_test_clean_tgmed/wer_12_0.0:%WER 4.90 [ 2575 / 52576, 345 ins, 229 del, 2001 sub ]
exp/chain/tdnn7m23t_sp/decode_test_clean_tgsmall/wer_12_0.0:%WER 5.35 [ 2811 / 52576, 340 ins, 269 del, 2202 sub ]
exp/chain/tdnn7m23t_sp/decode_test_other_fglarge/wer_13_0.5:%WER 8.97 [ 4694 / 52343, 564 ins, 481 del, 3649 sub ]
exp/chain/tdnn7m23t_sp/decode_test_other_tglarge/wer_13_0.5:%WER 9.42 [ 4931 / 52343, 581 ins, 542 del, 3808 sub ]
exp/chain/tdnn7m23t_sp/decode_test_other_tgmed/wer_13_0.0:%WER 11.56 [ 6053 / 52343, 690 ins, 659 del, 4704 sub ]
exp/chain/tdnn7m23t_sp/decode_test_other_tgsmall/wer_13_0.0:%WER 12.64 [ 6615 / 52343, 678 ins, 791 del, 5146 sub ]

which is consistently better than our earlier tdnn_lstm setup (which was adapted from mini-librispeech), for which we recorded:

exp/chain_cleaned/tdnn_lstm1a_sp/decode_dev_clean_fglarge/wer_10_0.5:%WER 3.64 [ 1980 / 54402, 232 ins, 170 del, 1578 sub ]
exp/chain_cleaned/tdnn_lstm1a_sp/decode_dev_clean_tglarge/wer_11_0.0:%WER 3.74 [ 2037 / 54402, 241 ins, 166 del, 1630 sub ]
exp/chain_cleaned/tdnn_lstm1a_sp/decode_dev_clean_tgmed/wer_10_0.0:%WER 4.70 [ 2558 / 54402, 266 ins, 236 del, 2056 sub ]
exp/chain_cleaned/tdnn_lstm1a_sp/decode_dev_clean_tgsmall/wer_10_0.0:%WER 5.11 [ 2780 / 54402, 266 ins, 274 del, 2240 sub ]
exp/chain_cleaned/tdnn_lstm1a_sp/decode_dev_other_fglarge/wer_12_0.0:%WER 10.01 [ 5100 / 50948, 634 ins, 474 del, 3992 sub ]
exp/chain_cleaned/tdnn_lstm1a_sp/decode_dev_other_tglarge/wer_13_0.0:%WER 10.56 [ 5381 / 50948, 626 ins, 584 del, 4171 sub ]
exp/chain_cleaned/tdnn_lstm1a_sp/decode_dev_other_tgmed/wer_13_0.0:%WER 12.56 [ 6401 / 50948, 652 ins, 818 del, 4931 sub ]
exp/chain_cleaned/tdnn_lstm1a_sp/decode_dev_other_tgsmall/wer_12_0.0:%WER 13.57 [ 6916 / 50948, 669 ins, 877 del, 5370 sub ]
exp/chain_cleaned/tdnn_lstm1a_sp/decode_looped_dev_clean_tgsmall/wer_11_0.0:%WER 5.08 [ 2766 / 54402, 264 ins, 281 del, 2221 sub ]
exp/chain_cleaned/tdnn_lstm1a_sp/decode_looped_dev_other_tgsmall/wer_14_0.0:%WER 13.92 [ 7092 / 50948, 634 ins, 1009 del, 5449 sub ]
exp/chain_cleaned/tdnn_lstm1a_sp/decode_looped_test_clean_tgsmall/wer_12_0.0:%WER 5.55 [ 2919 / 52576, 326 ins, 290 del, 2303 sub ]
exp/chain_cleaned/tdnn_lstm1a_sp/decode_looped_test_other_tgsmall/wer_14_0.0:%WER 14.03 [ 7344 / 52343, 635 ins, 1068 del, 5641 sub ]
exp/chain_cleaned/tdnn_lstm1a_sp/decode_test_clean_fglarge/wer_10_0.5:%WER 4.00 [ 2105 / 52576, 286 ins, 176 del, 1643 sub ]
exp/chain_cleaned/tdnn_lstm1a_sp/decode_test_clean_tglarge/wer_10_0.5:%WER 4.13 [ 2174 / 52576, 286 ins, 196 del, 1692 sub ]
exp/chain_cleaned/tdnn_lstm1a_sp/decode_test_clean_tgmed/wer_11_0.5:%WER 5.06 [ 2661 / 52576, 272 ins, 288 del, 2101 sub ]
exp/chain_cleaned/tdnn_lstm1a_sp/decode_test_clean_tgsmall/wer_10_0.0:%WER 5.53 [ 2907 / 52576, 333 ins, 272 del, 2302 sub ]
exp/chain_cleaned/tdnn_lstm1a_sp/decode_test_other_fglarge/wer_12_0.5:%WER 10.22 [ 5349 / 52343, 557 ins, 654 del, 4138 sub ]
exp/chain_cleaned/tdnn_lstm1a_sp/decode_test_other_tglarge/wer_11_0.5:%WER 10.78 [ 5641 / 52343, 580 ins, 640 del, 4421 sub ]
exp/chain_cleaned/tdnn_lstm1a_sp/decode_test_other_tgmed/wer_12_0.0:%WER 12.81 [ 6707 / 52343, 652 ins, 829 del, 5226 sub ]
exp/chain_cleaned/tdnn_lstm1a_sp/decode_test_other_tgsmall/wer_12_0.0:%WER 13.85 [ 7247 / 52343, 650 ins, 988 del, 5609 sub ]