Commit 8063c4ff authored by Dan Povey's avatar Dan Povey
Browse files

Updates to WSJ results

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@638 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8
parent ef9a515d
...@@ -65,12 +65,20 @@ exp/tri4b_mmi_b0.1/decode_tgpr_dev93/wer_15:%WER 11.82 [ 973 / 8234, 204 ins, 86 ...@@ -65,12 +65,20 @@ exp/tri4b_mmi_b0.1/decode_tgpr_dev93/wer_15:%WER 11.82 [ 973 / 8234, 204 ins, 86
# LDA+MLLT+SAT, SI-284, full retraining starting from 3b [c.f. 4b] # LDA+MLLT+SAT, SI-284, full retraining starting from 3b [c.f. 4b]
exp/tri4c/decode_tgpr_dev93/wer_16:%WER 12.92 [ 1064 / 8234, 224 ins, 94 del, 746 sub ] exp/tri4c/decode_tgpr_dev93/wer_16:%WER 12.92 [ 1064 / 8234, 224 ins, 94 del, 746 sub ]
# Mixing up further:
exp/tri4c_50k/decode_tgpr_dev93/wer_16:%WER 12.67 [ 1043 / 8234, 218 ins, 94 del, 731 sub ]
exp/tri4c_75k/decode_tgpr_dev93/wer_13:%WER 12.10 [ 996 / 8234, 225 ins, 82 del, 689 sub ]
exp/tri4c_100k/decode_tgpr_dev93/wer_15:%WER 11.89 [ 979 / 8234, 211 ins, 87 del, 681 sub ]
# sgmm4b is LDA+MLLT+SAT, on just SI-84 data. # sgmm4b is LDA+MLLT+SAT, on just SI-84 data.
## broken results, need to fix: [RE transitions] exp/sgmm4b/decode_tgpr_dev93/wer_14:%WER 12.69 [ 1045 / 8234, 204 ins, 104 del, 737 sub ]
#exp/sgmm4b/decode_tgpr_dev93/wer_14:%WER 13.29 [ 1094 / 8234, 213 ins, 124 del, 757 sub ] exp/sgmm4b/decode_tgpr_eval92/wer_11:%WER 8.63 [ 487 / 5643, 125 ins, 26 del, 336 sub ]
#exp/sgmm4b/decode_tgpr_eval92/wer_12:%WER 8.79 [ 496 / 5643, 122 ins, 30 del, 344 sub ] # mixing up a bit more:
exp/sgmm4b_12500/decode_tgpr_eval92/wer_11:%WER 8.56 [ 483 / 5643, 122 ins, 25 del, 336 sub ]
exp/sgmm4b_15000/decode_tgpr_eval92/wer_13:%WER 8.72 [ 492 / 5643, 120 ins, 28 del, 344 sub ]
# increasing subspace dim to 50.
exp/sgmm4b_50/decode_tgpr_eval92/wer_14:%WER 8.45 [ 477 / 5643, 110 ins, 27 del, 340 sub ]
# sgmm4c is the same, but on all SI-284 data. # sgmm4c is the same, but on all SI-284 data.
exp/sgmm4c/decode_tgpr_dev93/wer_11:%WER 10.74 [ 884 / 8234, 183 ins, 79 del, 622 sub ] exp/sgmm4c/decode_tgpr_dev93/wer_11:%WER 10.74 [ 884 / 8234, 183 ins, 79 del, 622 sub ]
......
...@@ -289,8 +289,16 @@ scripts/decode.sh --cmd "$decode_cmd" steps/decode_lda_mllt_sat.sh exp/tri4c/gra ...@@ -289,8 +289,16 @@ scripts/decode.sh --cmd "$decode_cmd" steps/decode_lda_mllt_sat.sh exp/tri4c/gra
75000 data/train_si284 exp/tri4c_50k exp/tri3b_ali_si284_20 exp/tri4c_75k 75000 data/train_si284 exp/tri4c_50k exp/tri3b_ali_si284_20 exp/tri4c_75k
scripts/decode.sh --cmd "$decode_cmd" steps/decode_lda_mllt_sat.sh exp/tri4c/graph_tgpr \ scripts/decode.sh --cmd "$decode_cmd" steps/decode_lda_mllt_sat.sh exp/tri4c/graph_tgpr \
data/test_dev93 exp/tri4c_75k/decode_tgpr_dev93 data/test_dev93 exp/tri4c_75k/decode_tgpr_dev93
steps/mixup_lda_etc.sh --num-jobs 20 --cmd "$train_cmd" \
100000 data/train_si284 exp/tri4c_75k exp/tri3b_ali_si284_20 exp/tri4c_100k
scripts/decode.sh --cmd "$decode_cmd" steps/decode_lda_mllt_sat.sh exp/tri4c/graph_tgpr \
data/test_dev93 exp/tri4c_100k/decode_tgpr_dev93
) )
# Train SGMM on top of LDA+MLLT+SAT, on all SI-284 data. C.f. 4b which was # Train SGMM on top of LDA+MLLT+SAT, on all SI-284 data. C.f. 4b which was
# just on SI-84. # just on SI-84.
steps/train_ubm_lda_etc.sh --num-jobs 20 --cmd "$train_cmd" \ steps/train_ubm_lda_etc.sh --num-jobs 20 --cmd "$train_cmd" \
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment