Merge pull request #4847 from sendream/master

Add recipe of Tibetan Amdo dialect

egs/xbmu_amdo31/README.txt

@@ -0,0 +1,11 @@
+About the XBMU-AMDO31 corpus XBMU-AMDO31 is an open-source Amdo Tibetan speech corpus published by Northwest Minzu University. 
+
+XBMU-AMDO31 dataset is a speech recognition corpus of Tibetan Amdo dialect. The open source corpus contains 31 hours of speech data and resources related to build speech recognition systems,including transcribed texts and a Tibetan pronunciation lexicon. (The lexicon is a Tibetan lexicon of the Lhasa dialect, which has been reused for the Amdo dialect because of the uniformity of the Tibetan language) The dataset can be used to train a model for Amdo Tibetan Automatic Speech Recognition (ASR).
+
+The database can be downloaded from openslr:
+http://www.openslr.org/133/
+
+For more details, please visit: 
+https://huggingface.co/datasets/syzym/xbmu_amdo31
+
+This recipe includes some different ASR models trained with XBMU-AMDO31.

egs/xbmu_amdo31/s5/RESULTS

@@ -0,0 +1,8 @@
+%WER 46.16 [ 15522 / 33628, 380 ins, 2208 del, 12934 sub ] exp/mono/decode_test/wer_10_0.0
+%WER 24.60 [ 8274 / 33628, 330 ins, 860 del, 7084 sub ] exp/tri1/decode_test/wer_13_0.0
+%WER 24.42 [ 8213 / 33628, 323 ins, 847 del, 7043 sub ] exp/tri2/decode_test/wer_13_0.0
+%WER 22.93 [ 7712 / 33628, 336 ins, 814 del, 6562 sub ] exp/tri3a/decode_test/wer_12_0.0
+%WER 20.17 [ 6783 / 33628, 275 ins, 764 del, 5744 sub ] exp/tri4a/decode_test/wer_15_0.0
+%WER 19.03 [ 6400 / 33628, 292 ins, 667 del, 5441 sub ] exp/tri5a/decode_test/wer_14_0.0
+%WER 15.45 [ 5196 / 33628, 229 ins, 646 del, 4321 sub ] exp/nnet3/tdnn_sp/decode_test/wer_16_0.0
+%WER 15.57 [ 5235 / 33628, 244 ins, 575 del, 4416 sub ] exp/chain/tdnn_1a_sp/decode_test/wer_11_0.0

egs/xbmu_amdo31/s5/cmd.sh

@@ -0,0 +1,15 @@
+# you can change cmd.sh depending on what type of queue you are using.
+# If you have no queueing system and want to run on a local machine, you
+# can change all instances 'queue.pl' to run.pl (but be careful and run
+# commands one by one: most recipes will exhaust the memory on your
+# machine).  queue.pl works with GridEngine (qsub).  slurm.pl works
+# with slurm.  Different queues are configured differently, with different
+# queue names and different ways of specifying things like memory;
+# to account for these differences you can create and edit the file
+# conf/queue.conf to match your queue's configuration.  Search for
+# conf/queue.conf in http://kaldi-asr.org/doc/queue.html for more information,
+# or search for the string 'default_config' in utils/queue.pl or utils/slurm.pl.
+
+export train_cmd="queue.pl --mem 2G"
+export decode_cmd="queue.pl --mem 4G"
+export mkgraph_cmd="queue.pl --mem 8G"

egs/xbmu_amdo31/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/xbmu_amdo31/s5/conf/mfcc.conf

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

egs/xbmu_amdo31/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/xbmu_amdo31/s5/conf/online_cmvn.conf

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

egs/xbmu_amdo31/s5/conf/online_pitch.conf

@@ -0,0 +1,4 @@
+--sample-frequency=16000
+--simulate-first-pass-online=true
+--normalization-right-context=25
+--frames-per-chunk=10

egs/xbmu_amdo31/s5/conf/pitch.conf

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

egs/xbmu_amdo31/s5/local/chain/run_tdnn.sh

@@ -0,0 +1 @@
+tuning/run_tdnn_1a.sh

egs/xbmu_amdo31/s5/local/chain/tuning/run_tdnn_1a.sh

@@ -0,0 +1,184 @@
+#!/usr/bin/env bash
+
+# This script is based on run_tdnn_7h.sh in swbd chain recipe.
+
+set -e
+
+# configs for 'chain'
+affix=
+stage=0
+train_stage=-10
+get_egs_stage=-10
+dir=exp/chain/tdnn_1a  # 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=1
+num_jobs_final=2
+minibatch_size=128
+frames_per_eg=150,110,90
+remove_egs=true
+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
+  if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $dir/egs/storage ]; then
+    utils/create_split_dir.pl \
+     /export/b0{5,6,7,8}/$USER/kaldi-data/egs/aishell-$(date +'%m_%d_%H_%M')/s5c/$dir/egs/storage $dir/egs/storage
+  fi
+
+  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 \
+    --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 [ $stage -le 13 ]; then
+  for test_set in dev test; do
+    steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
+      --nj 5 --cmd "$decode_cmd" \
+      --online-ivector-dir exp/nnet3/ivectors_$test_set \
+      $graph_dir data/${test_set}_hires $dir/decode_${test_set} || exit 1;
+  done
+fi
+
+exit;