Nara_WPE

egs/chime5/s5/cmd.sh

@@ -10,6 +10,6 @@
 # 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="retry.pl queue.pl --mem 2G"
+export train_cmd="retry.pl queue.pl --mem 120G"
 export decode_cmd="queue.pl --mem 4G"
 

egs/chime5/s5/local/chain/tuning/run_tdnn_1a.sh

@@ -198,10 +198,6 @@ if [ $stage -le 14 ]; then
     --trainer.num-chunk-per-minibatch=256,128,64 \
     --trainer.optimization.momentum=0.0 \
     --egs.chunk-width=$chunk_width \
-    --egs.chunk-left-context=$chunk_left_context \
-    --egs.chunk-right-context=$chunk_right_context \
-    --egs.chunk-left-context-initial=0 \
-    --egs.chunk-right-context-final=0 \
     --egs.dir="$common_egs_dir" \
     --egs.opts="--frames-overlap-per-eg 0" \
     --cleanup.remove-egs=$remove_egs \

egs/chime5/s5/local/run_wpe.py

@@ -0,0 +1,54 @@
+#!/usr/bin/env python
+# Copyright 2018 Johns Hopkins University (Author: Aswin Shanmugam Subramanian)
+# Apache 2.0
+# Works with both python2 and python3
+
+import numpy as np
+import soundfile as sf
+import time
+import os, errno
+from tqdm import tqdm
+import argparse
+
+from nara_wpe.wpe import wpe
+from nara_wpe.utils import stft, istft 
+from nara_wpe import project_root
+
+parser = argparse.ArgumentParser()
+parser.add_argument('--files', '-f', nargs='+')
+args = parser.parse_args()
+
+input_files = args.files[:len(args.files)//2]
+output_files = args.files[len(args.files)//2:]
+out_dir = os.path.dirname(output_files[0])
+try: 
+    os.makedirs(out_dir)
+except OSError as e:
+    if e.errno != errno.EEXIST:
+        raise
+
+stft_options = dict(
+    size=512,
+    shift=128,
+    window_length=None,
+    fading=True,
+    pad=True,
+    symmetric_window=False
+)
+
+sampling_rate = 16000
+delay = 3
+iterations = 5
+taps = 10
+
+signal_list = [
+    sf.read(f)[0]
+    for f in input_files
+]
+y = np.stack(signal_list, axis=0)
+Y = stft(y, **stft_options).transpose(2, 0, 1)
+Z = wpe(Y, iterations=iterations, statistics_mode='full').transpose(1, 2, 0)
+z = istft(Z, size=stft_options['size'], shift=stft_options['shift'])
+
+for d in range(len(signal_list)):
+    sf.write(output_files[d], z[d,:], sampling_rate)

egs/chime5/s5/local/run_wpe.sh

@@ -0,0 +1,152 @@
+#!/bin/bash
+# Copyright 2018 Johns Hopkins University (Author: Aswin Shanmugam Subramanian)
+# Apache 2.0
+
+. ./cmd.sh
+. ./path.sh
+
+# Config:
+nj=8
+cmd=run.pl
+
+. utils/parse_options.sh || exit 1;
+
+if [ $# != 3 ]; then
+   echo "Wrong #arguments ($#, expected 3)"
+   echo "Usage: local/run_wpe.sh [options] <wav-in-dir> <wav-out-dir> <array-id>"
+   echo "main options (for others, see top of script file)"
+   echo "  --cmd <cmd>                              # Command to run in parallel with"
+   echo "  --nj 50                        # number of jobs for parallel processing"
+   exit 1;
+fi
+
+sdir=$1
+odir=$2
+array=$3
+task=`basename $sdir`
+expdir=exp/wpe/${task}_${array}
+# Set bash to 'debug' mode, it will exit on :
+# -e 'error', -u 'undefined variable', -o ... 'error in pipeline', -x 'print commands',
+set -e
+set -u
+set -o pipefail
+
+miniconda_dir=$HOME/miniconda3/
+if [ ! -d $miniconda_dir ]; then
+    echo "$miniconda_dir does not exist. Please run '../../../tools/extras/install_miniconda.sh' and '../../../tools/extras/install_wpe.sh';"
+fi
+
+# check if WPE is installed
+result=`$HOME/miniconda3/bin/python -c "\
+try:
+    import nara_wpe
+    print('1')
+except ImportError:
+    print('0')"`
+
+if [ "$result" == "1" ]; then
+    echo "WPE is installed"
+else
+    echo "WPE is not installed. Please run ../../../tools/extras/install_wpe.sh"
+    exit 1
+fi
+
+mkdir -p $odir
+mkdir -p $expdir/log
+
+# wavfiles.list can be used as the name of the output files
+output_wavfiles=$expdir/wavfiles.list
+find -L ${sdir} | grep -i ${array} > $expdir/channels_input
+cat $expdir/channels_input | awk -F '/' '{print $NF}' | sed "s@S@$odir\/S@g" > $expdir/channels_output
+paste -d" " $expdir/channels_input $expdir/channels_output > $output_wavfiles
+
+# split the list for parallel processing
+split_wavfiles=""
+for n in `seq $nj`; do
+  split_wavfiles="$split_wavfiles $output_wavfiles.$n"
+done
+utils/split_scp.pl $output_wavfiles $split_wavfiles || exit 1;
+
+echo -e "Dereverberation - $task - $array\n"
+# making a shell script for each job
+for n in `seq $nj`; do
+cat <<-EOF > $expdir/log/wpe.$n.sh
+while read line; do
+  $HOME/miniconda3/bin/python local/run_wpe.py \
+    --file \$line
+done < $output_wavfiles.$n
+EOF
+done
+
+chmod a+x $expdir/log/wpe.*.sh
+$cmd JOB=1:$nj $expdir/log/wpe.JOB.log \
+  $expdir/log/wpe.JOB.sh
+
+# #################
+# for task in dt et; do
+    # for nch in 1 2 8; do
+        # wdir=exp/wpe_${task}_${nch}ch
+        # mkdir -p $wdir/log
+        # arrays=$wdir/channels
+        # output_wavfiles=$wdir/wavfiles.list
+        # if [ ${nch} == 1 ]; then
+            # allwavs=`cat ${dir}/${task}_real_1ch_wav.scp | cut -d " " -f2`
+            # allwavs_output=`cat ${dir}/${task}_real_1ch_wpe_wav.scp | cut -d " " -f2`
+            # echo $allwavs | tr ' ' '\n' > $wdir/channels_input
+            # echo $allwavs_output | tr ' ' '\n' > $wdir/channels_output
+            # paste -d" " $wdir/channels_input $wdir/channels_output > $arrays
+        # elif [ ${nch} == 2 ]; then
+            # allwavs=`cat ${dir}/${task}_real_2ch_wav.scp | cut -d " " -f2`
+            # allwavs_output=`cat ${dir}/${task}_real_2ch_wpe_wav.scp | cut -d " " -f2`
+            # echo $allwavs | tr ' ' '\n' | rev | sort | rev | awk 'NR%2==1' > $wdir/channels.1st
+            # echo $allwavs | tr ' ' '\n' | rev | sort | rev | awk 'NR%2==0' > $wdir/channels.2nd
+            # echo $allwavs_output | tr ' ' '\n' | rev | sort | rev | awk 'NR%2==1' > $wdir/channels_output.1st
+            # echo $allwavs_output | tr ' ' '\n' | rev | sort | rev | awk 'NR%2==0' > $wdir/channels_output.2nd
+            # paste -d" " $wdir/channels.1st $wdir/channels.2nd $wdir/channels_output.1st $wdir/channels_output.2nd > $arrays
+        # elif [ ${nch} == 8 ]; then
+            # allwavs=`cat ${dir}/${task}_real_8ch_wav.scp | cut -d " " -f2`
+            # allwavs_output=`cat ${dir}/${task}_real_8ch_wpe_wav.scp | cut -d " " -f2`
+            # echo $allwavs | tr ' ' '\n' | rev | sort | rev | awk 'NR%8==1' > $wdir/channels.1st
+            # echo $allwavs | tr ' ' '\n' | rev | sort | rev | awk 'NR%8==2' > $wdir/channels.2nd
+            # echo $allwavs | tr ' ' '\n' | rev | sort | rev | awk 'NR%8==3' > $wdir/channels.3rd
+            # echo $allwavs | tr ' ' '\n' | rev | sort | rev | awk 'NR%8==4' > $wdir/channels.4th
+            # echo $allwavs | tr ' ' '\n' | rev | sort | rev | awk 'NR%8==5' > $wdir/channels.5th
+            # echo $allwavs | tr ' ' '\n' | rev | sort | rev | awk 'NR%8==6' > $wdir/channels.6th
+            # echo $allwavs | tr ' ' '\n' | rev | sort | rev | awk 'NR%8==7' > $wdir/channels.7th
+            # echo $allwavs | tr ' ' '\n' | rev | sort | rev | awk 'NR%8==0' > $wdir/channels.8th
+            # echo $allwavs_output | tr ' ' '\n' | rev | sort | rev | awk 'NR%8==1' > $wdir/channels_output.1st
+            # echo $allwavs_output | tr ' ' '\n' | rev | sort | rev | awk 'NR%8==2' > $wdir/channels_output.2nd
+            # echo $allwavs_output | tr ' ' '\n' | rev | sort | rev | awk 'NR%8==3' > $wdir/channels_output.3rd
+            # echo $allwavs_output | tr ' ' '\n' | rev | sort | rev | awk 'NR%8==4' > $wdir/channels_output.4th
+            # echo $allwavs_output | tr ' ' '\n' | rev | sort | rev | awk 'NR%8==5' > $wdir/channels_output.5th
+            # echo $allwavs_output | tr ' ' '\n' | rev | sort | rev | awk 'NR%8==6' > $wdir/channels_output.6th
+            # echo $allwavs_output | tr ' ' '\n' | rev | sort | rev | awk 'NR%8==7' > $wdir/channels_output.7th
+            # echo $allwavs_output | tr ' ' '\n' | rev | sort | rev | awk 'NR%8==0' > $wdir/channels_output.8th
+            # paste -d" " $wdir/channels.1st $wdir/channels.2nd $wdir/channels.3rd $wdir/channels.4th $wdir/channels.5th $wdir/channels.6th $wdir/channels.7th $wdir/channels.8th $wdir/channels_output.1st $wdir/channels_output.2nd $wdir/channels_output.3rd $wdir/channels_output.4th $wdir/channels_output.5th $wdir/channels_output.6th $wdir/channels_output.7th $wdir/channels_output.8th > $arrays
+        # fi
+
+        # # split the list for parallel processing
+        # split_wavfiles=""
+        # for n in `seq $nj`; do
+            # split_wavfiles="$split_wavfiles $output_wavfiles.$n"
+        # done
+        # utils/split_scp.pl $arrays $split_wavfiles || exit 1;
+
+        # echo -e "Dereverberation - $task - real - $nch ch\n"
+        # # making a shell script for each job
+	# for n in `seq $nj`; do
+	# cat <<-EOF > $wdir/log/wpe.$n.sh
+	# while read line; do
+	  # $HOME/miniconda3/bin/python local/run_wpe.py \
+	    # --file \$line
+	# done < $output_wavfiles.$n
+	# EOF
+	# done
+
+        # chmod a+x $wdir/log/wpe.*.sh
+        # $cmd JOB=1:$nj $wdir/log/wpe.JOB.log \
+          # $wdir/log/wpe.JOB.sh
+    # done
+# done
+
+echo "`basename $0` Done."

egs/chime5/s5/run.sh

@@ -28,7 +28,7 @@ json_dir=${chime5_corpus}/transcriptions
 audio_dir=${chime5_corpus}/audio
 
 # training and test data
-train_set=train_worn_u100k
+train_set=train_worn_u400k
 test_sets="dev_worn dev_${enhancement}_ref eval_${enhancement}_ref"
 
 # This script also needs the phonetisaurus g2p, srilm, beamformit
@@ -73,18 +73,28 @@ if [ $stage -le 4 ]; then
   # Beamforming using reference arrays
   # enhanced WAV directory
   enhandir=enhan
+  dereverb_dir=${PWD}/wav/wpe/
   for dset in dev eval; do
     for mictype in u01 u02 u03 u04 u05 u06; do
-      local/run_beamformit.sh --cmd "$train_cmd" \
+      local/run_wpe.sh --cmd "$train_cmd" \
 			      ${audio_dir}/${dset} \
+			      ${dereverb_dir}/${dset} \
+			      ${mictype}
+    done
+  done
+
+  for dset in dev eval; do
+    for mictype in u01 u02 u03 u04 u05 u06; do
+      local/run_beamformit.sh --cmd "$train_cmd" \
+			      ${dereverb_dir}/${dset} \
 			      ${enhandir}/${dset}_${enhancement}_${mictype} \
 			      ${mictype}
     done
   done
 
   for dset in dev eval; do
     local/prepare_data.sh --mictype ref "$PWD/${enhandir}/${dset}_${enhancement}_u0*" \
-			  ${json_dir}/${dset} data/${dset}_${enhancement}_ref
+			  ${json_dir}/${dset} data/${dset}_${enhancement}_dereverb_ref
   done
 fi
 
@@ -99,8 +109,8 @@ if [ $stage -le 5 ]; then
   # randomly extract first 100k utterances from all mics
   # if you want to include more training data, you can increase the number of array mic utterances
   utils/combine_data.sh data/train_uall data/train_u01 data/train_u02 data/train_u04 data/train_u05 data/train_u06
-  utils/subset_data_dir.sh data/train_uall 100000 data/train_u100k
-  utils/combine_data.sh data/${train_set} data/train_worn data/train_u100k
+  utils/subset_data_dir.sh data/train_uall 400000 data/train_u400k
+  utils/combine_data.sh data/${train_set} data/train_worn data/train_u400k
 
   # only use left channel for worn mic recognition
   # you can use both left and right channels for training
@@ -242,15 +252,40 @@ if [ $stage -le 16 ]; then
 fi
 
 if [ $stage -le 17 ]; then
-  # chain TDNN
-  local/chain/run_tdnn.sh --nj ${nj} --train-set ${train_set}_cleaned --test-sets "$test_sets" --gmm tri3_cleaned --nnet3-affix _${train_set}_cleaned
+  rm -r data/train_worn_cleaned 2>/dev/null || true
+  utils/copy_data_dir.sh data/${train_set}_cleaned data/train_worn_cleaned
+
+  awk '{print $1}' data/train_worn/wav.scp > data/train_worn_cleaned/recos.tmp
+  utils/filter_scp.pl data/train_worn_cleaned/recos.tmp \
+    data/${train_set}_cleaned/wav.scp > data/train_worn_cleaned/wav.scp
+
+  utils/fix_data_dir.sh data/train_worn_cleaned
+
+  rm -r data/train_u400k_cleaned 2>/dev/null || true
+  utils/copy_data_dir.sh data/${train_set}_cleaned data/train_u400k_cleaned
+
+  utils/filter_scp.pl --exclude data/train_worn_cleaned/recos.tmp \
+    data/${train_set}_cleaned/wav.scp > data/train_u400k_cleaned/wav.scp
+
+  utils/fix_data_dir.sh data/train_u400k_cleaned
 fi
 
 if [ $stage -le 18 ]; then
+  # chain TDNN
+  #local/chain/tuning/run_tdnn_rvb_1b.sh --nj ${nj} --train-set ${train_set}_cleaned --test-sets "$test_sets" --gmm tri3_cleaned --nnet3-affix _${train_set}_cleaned
+  local/chain/tuning/run_tdnn_rvb_1b.sh --stage 16 --nj ${nj} \
+    --train-set-clean train_worn_cleaned \
+    --train-set-noisy train_u400k_cleaned \
+    --combined-train-set ${train_set}_cleaned \
+    --test-sets "$test_sets" \
+    --gmm tri3_cleaned --nnet3-affix _${train_set}_cleaned_rvb
+fi
+
+if [ $stage -le 19 ]; then
   # final scoring to get the official challenge result
   # please specify both dev and eval set directories so that the search parameters
   # (insertion penalty and language model weight) will be tuned using the dev set
   local/score_for_submit.sh \
-      --dev exp/chain_${train_set}_cleaned/tdnn1a_sp/decode_dev_${enhancement}_ref \
-      --eval exp/chain_${train_set}_cleaned/tdnn1a_sp/decode_eval_${enhancement}_ref
+      --dev exp/chain_${train_set}_cleaned_rvb/tdnn_rvb_1b_sp/decode_dev_${enhancement}_ref \
+      --eval exp/chain_${train_set}_cleaned_rvb/tdnn_rvb_1b_sp/decode_eval_${enhancement}_ref
 fi