[egs] Add "formosa_speech" recipe (Taiwanese Mandarin ASR) (#2474)

egs/formosa/README.txt

@@ -0,0 +1,22 @@
+### Welcome to the demo recipe of the Formosa Speech in the Wild (FSW) Project ###
+
+The language habits of Taiwanese people are different from other Mandarin speakers (both accents and cultures) [1]. Especially Tainwaese use tranditional Chinese characters, i.e., 繁體中文). To address this issue, a Taiwanese speech corpus collection project "Formosa Speech in the Wild (FSW)" was initiated in 2017 to improve the development of Taiwanese-specific speech recognition techniques.
+
+FSW corpus will be a large-scale database of real-Life/multi-gene Taiwanese Spontaneous speech collected and transcribed from various sources (radio, TV, open courses, etc.). To demostrate that this database is a reasonable data resource for Taiwanese spontaneous speech recognition research, a baseline recipe is provied here for everybody, especially students, to develop their own systems easily and quickly.
+
+This recipe is based on the "NER-Trs-Vol1" corpus (about 150 hours broadcast radio speech selected from FSW). For more details, please visit: 
+* Formosa Speech in the Wild (FSW) project (https://sites.google.com/speech.ntut.edu.tw/fsw)
+
+If you want to apply the NER-Trs-Vol1 corpus, please contact Yuan-Fu Liao (廖元甫) via "yfliao@mail.ntut.edu.tw". This corpus is only for non-commercial research/education use and will be distributed via our GitLab server in https://speech.nchc.org.tw.
+
+Any bug, errors, comments or suggestions are very welcomed.
+
+Yuan-Fu Liao (廖元甫)
+Associate Professor
+Department of electronic Engineering,
+National Taipei University of Technology
+http://www.ntut.edu.tw/~yfliao
+yfliao@mail.ntut.edu.tw
+
+............
+[1] The languages of Taiwan consist of several varieties of languages under families of the Austronesian languages and the Sino-Tibetan languages. Taiwanese Mandarin, Hokkien, Hakka and Formosan languages are used by 83.5%, 81.9%, 6.6% and 1.4% of the population respectively (2010). Given the prevalent use of Taiwanese Hokkien, the Mandarin spoken in Taiwan has been to a great extent influenced by it.

egs/formosa/s5/RESULTS

@@ -0,0 +1,43 @@
+#
+# Reference results
+#
+# Experimental settings:
+#
+# training set:	show CS, BG, DA, QG, SR, SY and WK,	in total 18977 utt., 1,088,948 words
+# test set:	show JZ, GJ, KX and YX,			in total  2112 utt.,   135,972 words
+# eval set:     show JX, TD and WJ,                     in total  2222 utt.,   104,648 words
+#
+# lexicon: 274,036 words
+# phones (IPA):  196 (tonal)
+#
+
+# WER: test
+
+%WER 61.32 [ 83373 / 135972, 5458 ins, 19156 del, 58759 sub ] exp/mono/decode_test/wer_11_0.0
+%WER 41.00 [ 55742 / 135972, 6725 ins, 12763 del, 36254 sub ] exp/tri1/decode_test/wer_15_0.0
+%WER 40.41 [ 54948 / 135972, 7366 ins, 11505 del, 36077 sub ] exp/tri2/decode_test/wer_14_0.0
+%WER 38.67 [ 52574 / 135972, 6855 ins, 11250 del, 34469 sub ] exp/tri3a/decode_test/wer_15_0.0
+%WER 35.70 [ 48546 / 135972, 7197 ins,  9717 del, 31632 sub ] exp/tri4a/decode_test/wer_17_0.0
+%WER 32.11 [ 43661 / 135972, 6112 ins, 10185 del, 27364 sub ] exp/tri5a/decode_test/wer_17_0.5
+%WER 31.36 [ 42639 / 135972, 6846 ins, 8860 del, 26933 sub ] exp/tri5a_cleaned/decode_test/wer_17_0.5
+%WER 24.43 [ 33218 / 135972, 5524 ins,  7583 del, 20111 sub ] exp/nnet3/tdnn_sp/decode_test/wer_12_0.0
+%WER 23.95 [ 32568 / 135972, 4457 ins, 10271 del, 17840 sub ] exp/chain/tdnn_1a_sp/decode_test/wer_10_0.0
+%WER 23.54 [ 32006 / 135972, 4717 ins, 8644 del, 18645 sub ] exp/chain/tdnn_1b_sp/decode_test/wer_10_0.0
+%WER 20.64 [ 28067 / 135972, 4434 ins, 7946 del, 15687 sub ] exp/chain/tdnn_1c_sp/decode_test/wer_11_0.0
+%WER 20.98 [ 28527 / 135972, 4706 ins, 7816 del, 16005 sub ] exp/chain/tdnn_1d_sp/decode_test/wer_10_0.0
+
+# CER: test
+
+%WER 54.09 [ 116688 / 215718, 4747 ins, 24510 del, 87431 sub ] exp/mono/decode_test/cer_10_0.0
+%WER 32.61 [  70336 / 215718, 5866 ins, 16282 del, 48188 sub ] exp/tri1/decode_test/cer_13_0.0
+%WER 32.10 [  69238 / 215718, 6186 ins, 15772 del, 47280 sub ] exp/tri2/decode_test/cer_13_0.0
+%WER 30.40 [  65583 / 215718, 6729 ins, 13115 del, 45739 sub ] exp/tri3a/decode_test/cer_12_0.0
+%WER 27.53 [  59389 / 215718, 6311 ins, 13008 del, 40070 sub ] exp/tri4a/decode_test/cer_15_0.0
+%WER 24.21 [  52232 / 215718, 6425 ins, 11543 del, 34264 sub ] exp/tri5a/decode_test/cer_15_0.0
+%WER 23.41 [ 50492 / 215718, 6645 ins, 10997 del, 32850 sub ] exp/tri5a_cleaned/decode_test/cer_17_0.0
+%WER 17.07 [  36829 / 215718, 4734 ins,  9938 del, 22157 sub ] exp/nnet3/tdnn_sp/decode_test/cer_12_0.0
+%WER 16.83 [ 36305 / 215718, 4772 ins, 10810 del, 20723 sub ] exp/chain/tdnn_1a_sp/decode_test/cer_9_0.0
+%WER 16.44 [ 35459 / 215718, 4216 ins, 11278 del, 19965 sub ] exp/chain/tdnn_1b_sp/decode_test/cer_10_0.0
+%WER 13.72 [ 29605 / 215718, 4678 ins, 8066 del, 16861 sub ] exp/chain/tdnn_1c_sp/decode_test/cer_10_0.0
+%WER 14.08 [ 30364 / 215718, 5182 ins, 7588 del, 17594 sub ] exp/chain/tdnn_1d_sp/decode_test/cer_9_0.0
+

egs/formosa/s5/cmd.sh

@@ -0,0 +1,27 @@
+# "queue.pl" uses qsub.  The options to it are
+# options to qsub.  If you have GridEngine installed,
+# change this to a queue you have access to.
+# Otherwise, use "run.pl", which will run jobs locally
+# (make sure your --num-jobs options are no more than
+# the number of cpus on your machine.
+
+# Run locally:
+#export train_cmd=run.pl
+#export decode_cmd=run.pl
+
+# JHU cluster (or most clusters using GridEngine, with a suitable
+# conf/queue.conf).
+export train_cmd="queue.pl"
+export decode_cmd="queue.pl --mem 4G"
+
+host=$(hostname -f)
+if [ ${host#*.} == "fit.vutbr.cz" ]; then
+  queue_conf=$HOME/queue_conf/default.conf # see example /homes/kazi/iveselyk/queue_conf/default.conf,
+  export train_cmd="queue.pl --config $queue_conf --mem 2G --matylda 0.2"
+  export decode_cmd="queue.pl --config $queue_conf --mem 3G --matylda 0.1"
+  export cuda_cmd="queue.pl --config $queue_conf --gpu 1 --mem 10G --tmp 40G"
+elif [ ${host#*.} == "cm.cluster" ]; then
+  # MARCC bluecrab cluster:
+  export train_cmd="slurm.pl --time 4:00:00 "
+  export decode_cmd="slurm.pl --mem 4G --time 4:00:00 "
+fi

egs/formosa/s5/conf/decode.config

@@ -0,0 +1,5 @@
+beam=11.0 # beam for decoding.  Was 13.0 in the scripts.
+first_beam=8.0 # beam for 1st-pass decoding in SAT.
+
+
+

egs/formosa/s5/conf/mfcc.conf

@@ -0,0 +1,2 @@
+--use-energy=false   # only non-default option.
+--sample-frequency=16000

egs/formosa/s5/conf/mfcc_hires.conf

@@ -0,0 +1,10 @@
+# config for high-resolution MFCC features, intended for neural network training.
+# Note: we keep all cepstra, so it has the same info as filterbank features,
+# but MFCC is more easily compressible (because less correlated) which is why
+# we prefer this method.
+--use-energy=false   # use average of log energy, not energy.
+--sample-frequency=16000 #  Switchboard is sampled at 8kHz
+--num-mel-bins=40     # similar to Google's setup.
+--num-ceps=40     # there is no dimensionality reduction.
+--low-freq=40    # low cutoff frequency for mel bins
+--high-freq=-200 # high cutoff frequently, relative to Nyquist of 8000 (=3800)

egs/formosa/s5/conf/online_cmvn.conf

@@ -0,0 +1 @@
+# configuration file for apply-cmvn-online, used when invoking online2-wav-nnet3-latgen-faster.

egs/formosa/s5/conf/pitch.conf

@@ -0,0 +1 @@
+--sample-frequency=16000

egs/formosa/s5/local/chain/run_tdnn.sh

@@ -0,0 +1 @@
+tuning/run_tdnn_1d.sh

egs/formosa/s5/local/chain/tuning/run_tdnn_1a.sh

@@ -0,0 +1,181 @@
+#!/bin/bash
+
+# This script is based on run_tdnn_7h.sh in swbd chain recipe.
+
+set -e
+
+# configs for 'chain'
+affix=1a
+stage=0
+train_stage=-10
+get_egs_stage=-10
+dir=exp/chain/tdnn  # Note: _sp will get added to this
+decode_iter=
+
+# training options
+num_epochs=4
+initial_effective_lrate=0.001
+final_effective_lrate=0.0001
+max_param_change=2.0
+final_layer_normalize_target=0.5
+num_jobs_initial=2
+num_jobs_final=12
+minibatch_size=128
+frames_per_eg=150,110,90
+remove_egs=false
+common_egs_dir=
+xent_regularize=0.1
+
+# End configuration section.
+echo "$0 $@"  # Print the command line for logging
+
+. ./cmd.sh
+. ./path.sh
+. ./utils/parse_options.sh
+
+if ! cuda-compiled; then
+  cat <<EOF && exit 1
+This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
+If you want to use GPUs (and have them), go to src/, and configure and make on a machine
+where "nvcc" is installed.
+EOF
+fi
+
+# The iVector-extraction and feature-dumping parts are the same as the standard
+# nnet3 setup, and you can skip them by setting "--stage 8" if you have already
+# run those things.
+
+dir=${dir}${affix:+_$affix}_sp
+train_set=train_sp
+ali_dir=exp/tri5a_sp_ali
+treedir=exp/chain/tri6_7d_tree_sp
+lang=data/lang_chain
+
+
+# if we are using the speed-perturbed data we need to generate
+# alignments for it.
+local/nnet3/run_ivector_common.sh --stage $stage || exit 1;
+
+if [ $stage -le 7 ]; then
+  # Get the alignments as lattices (gives the LF-MMI training more freedom).
+  # use the same num-jobs as the alignments
+  nj=$(cat $ali_dir/num_jobs) || exit 1;
+  steps/align_fmllr_lats.sh --nj $nj --cmd "$train_cmd" data/$train_set \
+    data/lang exp/tri5a exp/tri5a_sp_lats
+  rm exp/tri5a_sp_lats/fsts.*.gz # save space
+fi
+
+if [ $stage -le 8 ]; then
+  # Create a version of the lang/ directory that has one state per phone in the
+  # topo file. [note, it really has two states.. the first one is only repeated
+  # once, the second one has zero or more repeats.]
+  rm -rf $lang
+  cp -r data/lang $lang
+  silphonelist=$(cat $lang/phones/silence.csl) || exit 1;
+  nonsilphonelist=$(cat $lang/phones/nonsilence.csl) || exit 1;
+  # Use our special topology... note that later on may have to tune this
+  # topology.
+  steps/nnet3/chain/gen_topo.py $nonsilphonelist $silphonelist >$lang/topo
+fi
+
+if [ $stage -le 9 ]; then
+  # Build a tree using our new topology. This is the critically different
+  # step compared with other recipes.
+  steps/nnet3/chain/build_tree.sh --frame-subsampling-factor 3 \
+      --context-opts "--context-width=2 --central-position=1" \
+      --cmd "$train_cmd" 5000 data/$train_set $lang $ali_dir $treedir
+fi
+
+if [ $stage -le 10 ]; then
+  echo "$0: creating neural net configs using the xconfig parser";
+
+  num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
+  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
+
+  mkdir -p $dir/configs
+  cat <<EOF > $dir/configs/network.xconfig
+  input dim=100 name=ivector
+  input dim=43 name=input
+
+  # please note that it is important to have input layer with the name=input
+  # as the layer immediately preceding the fixed-affine-layer to enable
+  # the use of short notation for the descriptor
+  fixed-affine-layer name=lda input=Append(-1,0,1,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
+
+  # the first splicing is moved before the lda layer, so no splicing here
+  relu-batchnorm-layer name=tdnn1 dim=625
+  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=625
+  relu-batchnorm-layer name=tdnn3 input=Append(-1,0,1) dim=625
+  relu-batchnorm-layer name=tdnn4 input=Append(-3,0,3) dim=625
+  relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=625
+  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=625
+
+  ## adding the layers for chain branch
+  relu-batchnorm-layer name=prefinal-chain input=tdnn6 dim=625 target-rms=0.5
+  output-layer name=output include-log-softmax=false dim=$num_targets max-change=1.5
+
+  # adding the layers for xent branch
+  # This block prints the configs for a separate output that will be
+  # trained with a cross-entropy objective in the 'chain' models... this
+  # has the effect of regularizing the hidden parts of the model.  we use
+  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
+  # 0.5 / args.xent_regularize is suitable as it means the xent
+  # final-layer learns at a rate independent of the regularization
+  # constant; and the 0.5 was tuned so as to make the relative progress
+  # similar in the xent and regular final layers.
+  relu-batchnorm-layer name=prefinal-xent input=tdnn6 dim=625 target-rms=0.5
+  output-layer name=output-xent dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5
+
+EOF
+  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
+fi
+
+if [ $stage -le 11 ]; then
+  steps/nnet3/chain/train.py --stage $train_stage \
+    --cmd "$decode_cmd" \
+    --feat.online-ivector-dir exp/nnet3/ivectors_${train_set} \
+    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
+    --chain.xent-regularize $xent_regularize \
+    --chain.leaky-hmm-coefficient 0.1 \
+    --chain.l2-regularize 0.00005 \
+    --chain.apply-deriv-weights false \
+    --chain.lm-opts="--num-extra-lm-states=2000" \
+    --egs.dir "$common_egs_dir" \
+    --egs.stage $get_egs_stage \
+    --egs.opts "--frames-overlap-per-eg 0" \
+    --egs.chunk-width $frames_per_eg \
+    --trainer.num-chunk-per-minibatch $minibatch_size \
+    --trainer.frames-per-iter 1500000 \
+    --trainer.num-epochs $num_epochs \
+    --trainer.optimization.num-jobs-initial $num_jobs_initial \
+    --trainer.optimization.num-jobs-final $num_jobs_final \
+    --trainer.optimization.initial-effective-lrate $initial_effective_lrate \
+    --trainer.optimization.final-effective-lrate $final_effective_lrate \
+    --trainer.max-param-change $max_param_change \
+    --cleanup.remove-egs $remove_egs \
+    --feat-dir data/${train_set}_hires \
+    --tree-dir $treedir \
+    --lat-dir exp/tri5a_sp_lats \
+    --use-gpu wait \
+    --dir $dir  || exit 1;
+fi
+
+if [ $stage -le 12 ]; then
+  # Note: it might appear that this $lang directory is mismatched, and it is as
+  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
+  # the lang directory.
+  utils/mkgraph.sh --self-loop-scale 1.0 data/lang_test $dir $dir/graph
+fi
+
+graph_dir=$dir/graph
+if [ $stage -le 13 ]; then
+  for test_set in test eval; do
+    steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
+      --nj 10 --cmd "$decode_cmd" \
+      --online-ivector-dir exp/nnet3/ivectors_$test_set \
+      $graph_dir data/${test_set}_hires $dir/decode_${test_set} || exit 1;
+  done
+  wait;
+fi
+
+exit 0;