Children speech ASR recipe for cmu_kids and cslu_kids

egs/cmu_cslu_kids/README

@@ -0,0 +1,21 @@
+This is an ASR recipe for children speech using cmu_kids and cslu_kids.
+Both of the corpora can be found on LDC:
+    - cmu_kids : https://catalog.ldc.upenn.edu/LDC97S63
+    - cslu_kids: https://catalog.ldc.upenn.edu/LDC2007S18
+
+To run this recipe, you'll need a copy of both corpora:
+    ./run.sh --cmu_kids <path_to_cmu_corpus> --cslu_kids <path_to_cslu_corpus>
+
+By default, this recipe will download an LM pretrained on LibriSpeech from 
+lm_url=www.openslr.org/resources/11. If you already have a copy of this LM 
+and do not wish to redownload, you can specify the LM path using the --lm_src option:
+    ./run.sh --cmu_kids <path_to_cmu_corpus> --cslu_kids <path_to_cslu_corpus>\
+        --lm_src <path_to_librispeech_lm>
+
+This recipe will also download and clean CMU_Dict by default. If you have a clean copy 
+already, or wish to use your own dictionary, simply copy your version of the dict to 
+        data/local/dict
+
+To run extra features for triphone models or VLTN, set the following options true:
+    ./run.sh --cmu_kids <path_to_cmu_corpus> --cslu_kids <path_to_cslu_corpus>\
+        --vtln true --extra_features true

egs/cmu_cslu_kids/s5/cmd.sh

@@ -0,0 +1,23 @@
+# 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
+export decode_cmd="queue.pl --mem 2G"
+# the use of cuda_cmd is deprecated, used only in 'nnet1',
+export cuda_cmd="queue.pl --gpu 1"
+
+if [[ "$(hostname -f)" == "*.fit.vutbr.cz" ]]; then
+  queue_conf=$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"
+fi

egs/cmu_cslu_kids/s5/conf/decode.config

@@ -0,0 +1,4 @@
+# Use wider-than-normal decoding beams for RM.
+first_beam=16.0
+beam=20.0
+lattice_beam=10.0

egs/cmu_cslu_kids/s5/conf/decode_dnn.config

@@ -0,0 +1,8 @@
+# In RM, the optimal decode LMWT is in range 2..5, which is different from usual 10..15
+# (it is caused by using simple rule-based LM, instead of n-gram LM),
+scoring_opts="--min-lmwt 2 --max-lmwt 10"
+# Still, it is better to use --acwt 0.1, both for decoding and sMBR,
+acwt=0.1
+# For this small task we can afford to have large beams,
+beam=30.0 # beam for decoding.  Was 13.0 in the scripts.
+lattice_beam=18.0 # this has most effect on size of the lattices.

egs/cmu_cslu_kids/s5/conf/mfcc.conf

@@ -0,0 +1,2 @@
+--use-energy=false   # only non-default option.
+--allow_downsample=true

egs/cmu_cslu_kids/s5/conf/mfcc_hires.conf

@@ -0,0 +1,11 @@
+# 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.
+--num-mel-bins=40     # similar to Google's setup.
+--num-ceps=40     # there is no dimensionality reduction.
+--low-freq=20     # low cutoff frequency for mel bins... this is high-bandwidth data, so
+                  # there might be some information at the low end.
+--high-freq=-400 # high cutoff frequently, relative to Nyquist of 8000 (=7600)
+--allow-downsample=true

egs/cmu_cslu_kids/s5/conf/online_cmvn.conf

@@ -0,0 +1 @@
+# configuration file for apply-cmvn-online, used in the script ../local/run_online_decoding.sh

egs/cmu_cslu_kids/s5/conf/plp.conf

@@ -0,0 +1,2 @@
+# No non-default options for now.
+--allow_downsample=true 

egs/cmu_cslu_kids/s5/local/chain/compare_wer.sh

@@ -0,0 +1,137 @@
+#!/bin/bash
+
+# this script is used for comparing decoding results between systems.
+# e.g. local/chain/compare_wer.sh exp/chain/tdnn_{c,d}_sp
+# For use with discriminatively trained systems you specify the epochs after a colon:
+# for instance,
+# local/chain/compare_wer.sh exp/chain/tdnn_c_sp exp/chain/tdnn_c_sp_smbr:{1,2,3}
+
+
+if [ $# == 0 ]; then
+  echo "Usage: $0: [--looped] [--online] <dir1> [<dir2> ... ]"
+  echo "e.g.: $0 exp/chain/tdnn_{b,c}_sp"
+  echo "or (with epoch numbers for discriminative training):"
+  echo "$0 exp/chain/tdnn_b_sp_disc:{1,2,3}"
+  exit 1
+fi
+
+echo "# $0 $*"
+
+include_looped=false
+if [ "$1" == "--looped" ]; then
+  include_looped=true
+  shift
+fi
+include_online=false
+if [ "$1" == "--online" ]; then
+  include_online=true
+  shift
+fi
+
+
+used_epochs=false
+
+# this function set_names is used to separate the epoch-related parts of the name
+# [for discriminative training] and the regular parts of the name.
+# If called with a colon-free directory name, like:
+#  set_names exp/chain/tdnn_lstm1e_sp_bi_smbr
+# it will set dir=exp/chain/tdnn_lstm1e_sp_bi_smbr and epoch_infix=""
+# If called with something like:
+#  set_names exp/chain/tdnn_d_sp_smbr:3
+# it will set dir=exp/chain/tdnn_d_sp_smbr and epoch_infix="_epoch3"
+
+
+set_names() {
+  if [ $# != 1 ]; then
+    echo "compare_wer_general.sh: internal error"
+    exit 1  # exit the program
+  fi
+  dirname=$(echo $1 | cut -d: -f1)
+  epoch=$(echo $1 | cut -s -d: -f2)
+  if [ -z $epoch ]; then
+    epoch_infix=""
+  else
+    used_epochs=true
+    epoch_infix=_epoch${epoch}
+  fi
+}
+
+
+
+echo -n "# System               "
+for x in $*; do   printf "% 10s" " $(basename $x)";   done
+echo
+
+strings=(
+  "#WER dev_clean_2 (tgsmall) "
+  "#WER dev_clean_2 (tglarge) ")
+
+for n in 0 1; do
+   echo -n "${strings[$n]}"
+   for x in $*; do
+     set_names $x  # sets $dirname and $epoch_infix
+    decode_names=(tgsmall_dev_clean_2 tglarge_dev_clean_2)
+
+     wer=$(cat $dirname/decode_${decode_names[$n]}/wer_* | utils/best_wer.sh | awk '{print $2}')
+     printf "% 10s" $wer
+   done
+   echo
+   if $include_looped; then
+     echo -n "#             [looped:]    "
+     for x in $*; do
+       set_names $x  # sets $dirname and $epoch_infix
+       wer=$(cat $dirname/decode_looped_${decode_names[$n]}/wer_* | utils/best_wer.sh | awk '{print $2}')
+       printf "% 10s" $wer
+     done
+     echo
+   fi
+   if $include_online; then
+     echo -n "#             [online:]    "
+     for x in $*; do
+       set_names $x  # sets $dirname and $epoch_infix
+       wer=$(cat ${dirname}_online/decode_${decode_names[$n]}/wer_* | utils/best_wer.sh | awk '{print $2}')
+       printf "% 10s" $wer
+     done
+     echo
+   fi
+done
+
+
+if $used_epochs; then
+  exit 0;  # the diagnostics aren't comparable between regular and discriminatively trained systems.
+fi
+
+
+echo -n "# Final train prob     "
+for x in $*; do
+  prob=$(grep Overall $x/log/compute_prob_train.final.log | grep -v xent | awk '{printf("%.4f", $8)}')
+  printf "% 10s" $prob
+done
+echo
+
+echo -n "# Final valid prob     "
+for x in $*; do
+  prob=$(grep Overall $x/log/compute_prob_valid.final.log | grep -v xent | awk '{printf("%.4f", $8)}')
+  printf "% 10s" $prob
+done
+echo
+
+echo -n "# Final train prob (xent)"
+for x in $*; do
+  prob=$(grep Overall $x/log/compute_prob_train.final.log | grep -w xent | awk '{printf("%.4f", $8)}')
+  printf "% 10s" $prob
+done
+echo
+
+echo -n "# Final valid prob (xent)"
+for x in $*; do
+  prob=$(grep Overall $x/log/compute_prob_valid.final.log | grep -w xent | awk '{printf("%.4f", $8)}')
+  printf "% 10s" $prob
+done
+echo
+
+echo -n "# Num-params              "
+for x in $*; do
+  printf "% 10s" $(grep num-parameters $x/log/progress.1.log | awk '{print $2}')
+done
+echo

egs/cmu_cslu_kids/s5/local/chain/run_tdnnf.sh

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

egs/cmu_cslu_kids/s5/local/chain/tdnnf_decode.sh

@@ -0,0 +1,82 @@
+#! /bin/bash
+
+# Copyright Johns Hopkins University
+#   2019 Fei Wu
+
+# Decode on new data set using trained model. 
+# The data directory should be prepared in kaldi style.
+# Usage:
+#     ./local/chain/tdnnF_decode.sh --data_src <prepared_data_dir> 
+
+set -euo pipefail
+echo "$0 $@"
+
+stage=0
+decode_nj=10
+data_src=
+affix=
+tree_affix=
+nnet3_affix=
+
+. ./cmd.sh
+. ./path.sh
+. ./utils/parse_options.sh
+
+if ! cuda-compiled; then
+  cat <<EOF && exit 1
+This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
+If you want to use GPUs (and have them), go to src/, and configure and make on a machine
+where "nvcc" is installed.
+EOF
+fi
+data_name=$(basename $data_src)
+data_hires="data/${data_name}_hires"
+ivect_dir=exp/nnet3${nnet3_affix}/ivector_$data_name
+tree_dir=exp/chain${nnet3_affix}/tree_sp${tree_affix:+_$tree_affix}
+dir=exp/chain${nnet3_affix}/tdnn${affix}_sp
+
+mfcc=mfcc_hires_$data_name
+chunk_width=140,100,160
+reporting_email=
+
+if [ $stage -le 0 ]; then 
+    rm -rf $data_hires
+    cp -r $data_src $data_hires
+fi
+# High resolution mfcc
+if [ $stage -le 1 ]; then 
+    steps/make_mfcc.sh --nj 10 --mfcc-config conf/mfcc_hires.conf \
+        --cmd "$train_cmd" $data_hires \
+        exp/$data_name/make_feat_hires  $mfcc|| exit 1;
+    steps/compute_cmvn_stats.sh $data_hires || exit 1;
+    utils/fix_data_dir.sh $data_hires || exit 1;
+fi
+
+# Extract i-vector
+if [ $stage -le 2 ]; then 
+    steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj 20 \
+        $data_hires exp/nnet3"$affix"/extractor $ivect_dir
+fi
+
+if [ $stage -le 3 ]; then
+  frames_per_chunk=$(echo $chunk_width | cut -d, -f1)
+  rm $dir/.error 2>/dev/null || true
+
+    (
+      nspk=$(wc -l <$data_hires/spk2utt)
+      steps/nnet3/decode.sh \
+          --acwt 1.0 --post-decode-acwt 10.0 \
+          --frames-per-chunk $frames_per_chunk \
+          --nj $nspk --cmd "$decode_cmd"  --num-threads 4 \
+          --online-ivector-dir $ivect_dir \
+          $tree_dir/graph_tgsmall $data_hires ${dir}/decode_tgsmall_$data_name || exit 1
+
+      steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" \
+        data/lang_test_{tgsmall,tglarge} \
+       $data_hires ${dir}/decode_{tgsmall,tglarge}_$data_name || exit 1
+    ) || touch $dir/.error &
+
+  wait
+  [ -f $dir/.error ] && echo "$0: there was a problem while decoding" && exit 1
+fi
+