trunk: Adding DNN-based speaker recognition recipe in egs/sre10

git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@5223 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8

egs/sre08/v1/sid/extract_ivectors_dnn.sh

@@ -0,0 +1,94 @@
+#!/bin/bash
+
+# Copyright     2013  Daniel Povey
+#          2014-2015  David Snyder
+#               2015  Johns Hopkins University (Author: Daniel Garcia-Romero)
+#               2015  Johns Hopkins University (Author: Daniel Povey)
+# Apache 2.0.
+
+# This script extracts iVectors for a set of utterances, given
+# features and a trained DNN-based iVector extractor.
+
+# Begin configuration section.
+nj=30
+cmd="run.pl"
+stage=0
+min_post=0.025 # Minimum posterior to use (posteriors below this are pruned out)
+posterior_scale=1.0 # This scale helps to control for successive features being highly
+                    # correlated.  E.g. try 0.1 or 0.3.
+# End configuration section.
+
+echo "$0 $@"  # Print the command line for logging
+
+if [ -f path.sh ]; then . ./path.sh; fi
+. parse_options.sh || exit 1;
+
+
+if [ $# != 5 ]; then
+  echo "Usage: $0 <extractor-dir> <data> <ivector-dir>"
+  echo " e.g.: $0 exp/extractor_2048_male data/train_male exp/ivectors_male"
+  echo "main options (for others, see top of script file)"
+  echo "  --config <config-file>                           # config containing options"
+  echo "  --cmd (utils/run.pl|utils/queue.pl <queue opts>) # how to run jobs."
+  echo "  --num-iters <#iters|10>                          # Number of iterations of E-M"
+  echo "  --nj <n|10>                                      # Number of jobs (also see num-processes and num-threads)"
+  echo "  --num-threads <n|8>                              # Number of threads for each process"
+  echo "  --stage <stage|0>                                # To control partial reruns"
+  echo "  --num-gselect <n|20>                             # Number of Gaussians to select using"
+  echo "                                                   # diagonal model."
+  echo "  --min-post <min-post|0.025>                      # Pruning threshold for posteriors"
+  exit 1;
+fi
+
+srcdir=$1
+nnet=$2
+data=$3
+data_dnn=$4
+dir=$5
+
+for f in $srcdir/final.ie $srcdir/final.ubm $data/feats.scp ; do
+  [ ! -f $f ] && echo "No such file $f" && exit 1;
+done
+
+# Set various variables.
+mkdir -p $dir/log
+sdata=$data/split$nj;
+utils/split_data.sh $data $nj || exit 1;
+
+sdata_dnn=$data_dnn/split$nj;
+utils/split_data.sh $data_dnn $nj || exit 1;
+
+delta_opts=`cat $srcdir/delta_opts 2>/dev/null`
+
+splice_opts=`cat exp/nnet//splice_opts 2>/dev/null` # frame-splicing options           
+
+## Set up features.
+feats="ark,s,cs:add-deltas $delta_opts scp:$sdata/JOB/feats.scp ark:- | apply-cmvn-sliding --norm-vars=false --center=true --cmn-window=300 ark:- ark:- | select-voiced-frames ark:- scp,s,cs:$sdata/JOB/vad.scp ark:- |"
+
+nnet_feats="ark,s,cs:apply-cmvn-sliding --center=true scp:$sdata_dnn/JOB/feats.scp ark:- |"
+
+if [ $stage -le 0 ]; then
+  echo "$0: extracting iVectors"
+  $cmd JOB=1:$nj $dir/log/extract_ivectors.JOB.log \
+    nnet-am-compute --apply-log=true $nnet "$nnet_feats" ark:- \
+    \| select-voiced-frames ark:- scp,s,cs:$sdata/JOB/vad.scp ark:- \
+    \| logprob-to-post --min-post=$min_post ark:- ark:- \| \
+    scale-post ark:- $posterior_scale ark:- \| \
+    ivector-extract --verbose=2 $srcdir/final.ie "$feats" ark,s,cs:- \
+      ark,scp,t:$dir/ivector.JOB.ark,$dir/ivector.JOB.scp || exit 1;
+fi
+
+if [ $stage -le 1 ]; then
+  echo "$0: combining iVectors across jobs"
+  for j in $(seq $nj); do cat $dir/ivector.$j.scp; done >$dir/ivector.scp || exit 1;
+fi
+
+if [ $stage -le 2 ]; then
+  # Be careful here: the speaker-level iVectors are now length-normalized,
+  # even if they are otherwise the same as the utterance-level ones.
+  echo "$0: computing mean of iVectors for each speaker and length-normalizing"
+  $cmd $dir/log/speaker_mean.log \
+    ivector-normalize-length scp:$dir/ivector.scp  ark:- \| \
+    ivector-mean ark:$data/spk2utt ark:- ark:- ark,t:$dir/num_utts.ark \| \
+    ivector-normalize-length ark:- ark,scp:$dir/spk_ivector.ark,$dir/spk_ivector.scp || exit 1;
+fi

egs/sre08/v1/sid/init_full_ubm_from_dnn.sh

@@ -0,0 +1,79 @@
+#!/bin/bash
+# Copyright 2015   David Snyder
+#           2015   Johns Hopkins University (Author: Daniel Garcia-Romero)
+#           2015   Johns Hopkins University (Author: Daniel Povey)
+# Apache 2.0
+
+# This script derives a full-covariance UBM from DNN posteriors and
+# speaker recognition features.
+
+# Begin configuration section.
+nj=40
+cmd="run.pl"
+stage=-2
+delta_window=3
+delta_order=2
+num_components=5297
+# End configuration section.
+
+echo "$0 $@"  # Print the command line for logging
+
+if [ -f path.sh ]; then . ./path.sh; fi
+. parse_options.sh || exit 1;
+
+if [ $# != 3 ]; then
+  echo "Usage: steps/init_full_ubm_from_dnn.sh <data-speaker-id> <data-dnn> <dnn-model> <new-ubm-dir>"
+  echo "Initializes a full-covariance UBM from DNN posteriors and speaker recognition features."
+  echo " e.g.: steps/init_full_ubm_from_dnn.sh data/train data/train_dnn exp/dnn/final.mdl exp/full_ubm"
+  echo "main options (for others, see top of script file)"
+  echo "  --config <config-file>                           # config containing options"
+  echo "  --cmd (utils/run.pl|utils/queue.pl <queue opts>) # how to run jobs."
+  echo "  --nj <n|16>                                      # number of parallel training jobs"
+  echo "  --delta-window <n|3>                             # delta window size"
+  echo "  --delta-order <n|2>                              # delta order"
+  echo "  --number-components <n|5297>                     # number of components in the final GMM needs"
+  echo "                                                   # to be equal to the size of the DNN output layer."
+  exit 1;
+fi
+
+data=$1
+data_dnn=$2
+nnet=$3
+dir=$4
+
+for f in $data/feats.scp $data/vad.scp; do
+  [ ! -f $f ] && echo "No such file $f" && exit 1;
+done
+
+mkdir -p $dir/log
+echo $nj > $dir/num_jobs
+sdata=$data/split$nj;
+utils/split_data.sh $data $nj || exit 1;
+
+sdata_dnn=$data_dnn/split$nj;
+utils/split_data.sh $data_dnn $nj || exit 1;
+
+delta_opts="--delta-window=$delta_window --delta-order=$delta_order"
+echo $delta_opts > $dir/delta_opts
+
+logdir=$dir/log
+
+nnet_feats="ark,s,cs:apply-cmvn-sliding --center=true scp:$sdata_dnn/JOB/feats.scp ark:- |"
+
+feats="ark,s,cs:add-deltas $delta_opts scp:$sdata/JOB/feats.scp ark:- | \
+apply-cmvn-sliding --norm-vars=false --center=true --cmn-window=300 ark:- ark:- | \
+select-voiced-frames ark:- scp,s,cs:$sdata/JOB/vad.scp ark:- |"
+
+$cmd JOB=1:$nj $logdir/make_stats.JOB.log \
+  nnet-am-compute --apply-log=true $nnet "$nnet_feats" ark:- \
+  \| select-voiced-frames ark:- scp,s,cs:$sdata/JOB/vad.scp ark:- \
+  \| logprob-to-post ark:- ark:- \| \
+  fgmm-global-acc-stats-post ark:- $num_components "$feats" \
+  $dir/stats.JOB.acc || exit 1;
+
+$cmd $dir/log/init.log \
+  fgmm-global-init-from-accs --verbose=2 \
+  "fgmm-global-sum-accs - $dir/stats.*.acc |" $num_components \
+  $dir/final.ubm || exit 1;
+
+exit 0;

egs/sre08/v1/sid/train_ivector_extractor_dnn.sh

@@ -0,0 +1,181 @@
+#!/bin/bash
+
+# Copyright 2013  Daniel Povey
+#      2014-2015  David Snyder
+#           2015  Johns Hopkins University (Author: Daniel Garcia-Romero)
+#           2015  Johns Hopkins University (Author: Daniel Povey)
+# Apache 2.0.
+
+# This script trains the i-vector extractor using a DNN-based UBM. It also requires
+# an fGMM, usually created by the script sid/init_full_gmm_from_dnn.sh.
+# Note: there are 3 separate levels of parallelization: num_threads, num_processes, 
+# and num_jobs.  This may seem a bit excessive.  It has to do with minimizing 
+# memory usage and disk I/O, subject to various constraints.  The "num_threads" 
+# is how many threads a program uses; the "num_processes" is the number of separate
+# processes a single  job spawns, and then sums the accumulators in memory.
+# Our recommendation:
+#  - Set num_threads to the minimum of (4, or how many virtual cores your machine has).
+#    (because of needing to lock various global quantities, the program can't
+#    use many more than 4 threads with good CPU utilization).
+#  - Set num_processes to the number of virtual cores on each machine you have, divided by 
+#    num_threads.  E.g. 4, if you have 16 virtual cores.   If you're on a shared queue
+#    that's busy with other people's jobs, it may be wise to set it to rather less
+#    than this maximum though, or your jobs won't get scheduled.  And if memory is
+#    tight you need to be careful; in our normal setup, each process uses about 5G.
+#  - Set num_jobs to as many of the jobs (each using $num_threads * $num_processes CPUs)
+#    your queue will let you run at one time, but don't go much more than 10 or 20, or
+#    summing the accumulators will possibly get slow.  If you have a lot of data, you
+#    may want more jobs, though.
+
+# Begin configuration section.
+nj=10   # this is the number of separate queue jobs we run, but each one 
+        # contains num_processes sub-jobs.. the real number of threads we 
+        # run is nj * num_processes * num_threads, and the number of
+        # separate pieces of data is nj * num_processes.
+num_threads=4
+num_processes=4 # each job runs this many processes, each with --num-threads threads
+cmd="run.pl"
+stage=-4
+num_gselect=20 # Gaussian-selection using diagonal model: number of Gaussians to select
+ivector_dim=400 # dimension of the extracted i-vector
+use_weights=false # set to true to turn on the regression of log-weights on the ivector.
+num_iters=10
+min_post=0.025 # Minimum posterior to use (posteriors below this are pruned out)
+num_samples_for_weights=3 # smaller than the default for speed (relates to a sampling method)
+cleanup=true
+posterior_scale=1.0 # This scale helps to control for successve features being highly
+                    # correlated.  E.g. try 0.1 or 0.3
+sum_accs_opt=
+# End configuration section.
+
+echo "$0 $@"  # Print the command line for logging
+
+if [ -f path.sh ]; then . ./path.sh; fi
+. parse_options.sh || exit 1;
+
+
+if [ $# != 5 ]; then
+  echo "Usage: $0 <fgmm-model> <dnn-model> <data-speaker-id> <data-dnn> <extractor-dir>"
+  echo " e.g.: $0 exp/sup_ubm/final.ubm exp/dnn/final.mdl data/train data/train_dnn exp/extractor_male"
+  echo "main options (for others, see top of script file)"
+  echo "  --config <config-file>                           # config containing options"
+  echo "  --cmd (utils/run.pl|utils/queue.pl <queue opts>) # how to run jobs."
+  echo "  --num-iters <#iters|10>                          # Number of iterations of E-M"
+  echo "  --nj <n|10>                                      # Number of jobs (also see num-processes and num-threads)"
+  echo "  --num-processes <n|4>                            # Number of processes for each queue job (relates"
+  echo "                                                   # to summing accs in memory)"
+  echo "  --num-threads <n|4>                              # Number of threads for each process (can't be usefully"
+  echo "                                                   # increased much above 4)"
+  echo "  --stage <stage|-4>                               # To control partial reruns"
+  echo "  --num-gselect <n|20>                             # Number of Gaussians to select using"
+  echo "                                                   # diagonal model."
+  echo "  --sum-accs-opt <option|''>                       # Option e.g. '-l hostname=a15' to localize"
+  echo "                                                   # sum-accs process to nfs server."
+  exit 1;
+fi
+
+fgmm_model=$1
+nnet=$2
+data=$3
+data_dnn=$4
+dir=$5
+
+srcdir=$(dirname $fgmm_model)
+
+for f in $fgmm_model $data/feats.scp ; do
+  [ ! -f $f ] && echo "No such file $f" && exit 1;
+done
+
+# Set various variables.
+mkdir -p $dir/log
+nj_full=$[$nj*$num_processes]
+sdata=$data/split$nj_full;
+utils/split_data.sh $data $nj_full || exit 1;
+
+sdata_dnn=$data_dnn/split$nj_full;
+utils/split_data.sh $data_dnn $nj_full || exit 1;
+
+delta_opts=`cat $srcdir/delta_opts 2>/dev/null`
+if [ -f $srcdir/delta_opts ]; then
+  cp $srcdir/delta_opts $dir/ 2>/dev/null
+fi
+
+splice_opts=`cat exp/nnet//splice_opts 2>/dev/null` # frame-splicing options           
+
+parallel_opts="-pe smp $[$num_threads*$num_processes]"
+## Set up features.
+feats="ark,s,cs:add-deltas $delta_opts scp:$sdata/JOB/feats.scp ark:- | apply-cmvn-sliding --norm-vars=false --center=true --cmn-window=300 ark:- ark:- | select-voiced-frames ark:- scp,s,cs:$sdata/JOB/vad.scp ark:- |"
+
+nnet_feats="ark,s,cs:apply-cmvn-sliding --center=true scp:$sdata_dnn/JOB/feats.scp ark:- |"
+
+
+# Initialize the i-vector extractor using the FGMM input
+if [ $stage -le -2 ]; then
+  cp $fgmm_model $dir/final.ubm || exit 1;
+  $cmd $dir/log/convert.log \
+    fgmm-global-to-gmm $dir/final.ubm $dir/final.dubm || exit 1;
+  $cmd $dir/log/init.log \
+    ivector-extractor-init --ivector-dim=$ivector_dim --use-weights=$use_weights \
+     $dir/final.ubm $dir/0.ie || exit 1;
+fi 
+
+# Do Gaussian selection and posterior extracion
+
+if [ $stage -le -1 ]; then
+  echo $nj_full > $dir/num_jobs
+  echo "$0: doing DNN posterior computation"
+  $cmd JOB=1:$nj_full $dir/log/post.JOB.log \
+  nnet-am-compute --apply-log=true $nnet "$nnet_feats" ark:- \
+  \| select-voiced-frames ark:- scp,s,cs:$sdata/JOB/vad.scp ark:- \
+  \| logprob-to-post --min-post=$min_post ark,s,cs:- ark:- \| \
+  scale-post ark:- $posterior_scale "ark:|gzip -c >$dir/post.JOB.gz" || exit 1;
+
+else
+  if ! [ $nj_full -eq $(cat $dir/num_jobs) ]; then
+    echo "Num-jobs mismatch $nj_full versus $(cat $dir/num_jobs)"
+    exit 1
+  fi
+fi
+
+x=0
+while [ $x -lt $num_iters ]; do
+  if [ $stage -le $x ]; then
+    rm $dir/.error 2>/dev/null
+
+    Args=() # bash array of training commands for 1:nj, that put accs to stdout.
+    for j in $(seq $nj_full); do
+      Args[$j]=`echo "ivector-extractor-acc-stats --num-threads=$num_threads --num-samples-for-weights=$num_samples_for_weights $dir/$x.ie '$feats' 'ark,s,cs:gunzip -c $dir/post.JOB.gz|' -|" | sed s/JOB/$j/g`
+    done
+
+    echo "Accumulating stats (pass $x)"
+    for g in $(seq $nj); do
+      start=$[$num_processes*($g-1)+1]
+      $cmd $parallel_opts $dir/log/acc.$x.$g.log \
+        ivector-extractor-sum-accs --parallel=true "${Args[@]:$start:$num_processes}" \
+          $dir/acc.$x.$g || touch $dir/.error &
+    done
+    wait
+    [ -f $dir/.error ] && echo "Error accumulating stats on iteration $x" && exit 1;
+	accs=""
+	for j in $(seq $nj); do
+	  accs+="$dir/acc.$x.$j "
+	done
+	echo "Summing accs (pass $x)"
+	$cmd $sum_accs_opt $dir/log/sum_acc.$x.log \
+	  ivector-extractor-sum-accs $accs $dir/acc.$x || exit 1;
+    echo "Updating model (pass $x)"
+    nt=$[$num_threads*$num_processes] # use the same number of threads that
+                                      # each accumulation process uses, since we
+                                      # can be sure the queue will support this many.
+	$cmd -pe smp $nt $dir/log/update.$x.log \
+	  ivector-extractor-est --num-threads=$nt $dir/$x.ie $dir/acc.$x $dir/$[$x+1].ie || exit 1;
+	rm $dir/acc.$x.*
+    if $cleanup; then
+      rm $dir/acc.$x
+      # rm $dir/$x.ie
+    fi
+  fi
+  x=$[$x+1]
+done
+
+ln -s $x.ie $dir/final.ie

egs/sre10/v1/local/dnn/README

@@ -0,0 +1,5 @@
+This directory contains DNN scripts based on the nnet2 recipes found in
+the ASR examples (e.g., fisher_english). The scripts have been modified
+for speaker recognition purposes. Most of the scripts are lightly modified
+versions of those appearing in the steps or local directories of
+egs/fisher_english.