New chime-5 recipe

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="queue.pl --mem 2G"
+export train_cmd="retry.pl queue.pl --mem 2G"
 export decode_cmd="queue.pl --mem 4G"
 

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

@@ -24,21 +24,16 @@ decode_iter=
 # training options
 # training chunk-options
 chunk_width=140,100,160
-# we don't need extra left/right context for TDNN systems.
-chunk_left_context=0
-chunk_right_context=0
 common_egs_dir=
 xent_regularize=0.1
 
 # training options
 srand=0
 remove_egs=true
-reporting_email=
 
 #decode options
 test_online_decoding=false  # if true, it will run the last decoding stage.
 
-
 # End configuration section.
 echo "$0 $@"  # Print the command line for logging
 
@@ -176,7 +171,6 @@ EOF
   steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
 fi
 
-
 if [ $stage -le 14 ]; then
   if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $dir/egs/storage ]; then
     utils/create_split_dir.pl \
@@ -204,15 +198,10 @@ 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 \
     --use-gpu=true \
-    --reporting.email="$reporting_email" \
     --feat-dir=$train_data_dir \
     --tree-dir=$tree_dir \
     --lat-dir=$lat_dir \
@@ -235,10 +224,6 @@ if [ $stage -le 16 ]; then
     (
       steps/nnet3/decode.sh \
           --acwt 1.0 --post-decode-acwt 10.0 \
-          --extra-left-context $chunk_left_context \
-          --extra-right-context $chunk_right_context \
-          --extra-left-context-initial 0 \
-          --extra-right-context-final 0 \
           --frames-per-chunk $frames_per_chunk \
           --nj 8 --cmd "$decode_cmd"  --num-threads 4 \
           --online-ivector-dir exp/nnet3${nnet3_affix}/ivectors_${data}_hires \

egs/chime5/s5/local/nnet3/run_ivector_common.sh

@@ -23,7 +23,7 @@ nnet3_affix=_train_worn_u100k
 gmm_dir=exp/${gmm}
 ali_dir=exp/${gmm}_ali_${train_set}_sp
 
-for f in data/${train_set}/feats.scp ${gmm_dir}/final.mdl; do
+for f in data/${train_set}/utt2spk ${gmm_dir}/final.mdl; do
   if [ ! -f $f ]; then
     echo "$0: expected file $f to exist"
     exit 1

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 +1,85 @@
+#!/bin/bash
+# Copyright 2018 Johns Hopkins University (Author: Aswin Shanmugam Subramanian)
+# Apache 2.0
+
+. ./cmd.sh
+. ./path.sh
+
+# Config:
+nj=4
+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
+
+echo "`basename $0` Done."

egs/chime5/s5b/RESULTS

@@ -0,0 +1,33 @@
+
+# tri2
+%WER 76.40 [ 44985 / 58881, 3496 ins, 17652 del, 23837 sub ] exp/tri2/decode_dev_worn/wer_13_1.0
+%WER 93.56 [ 55091 / 58881, 2132 ins, 35555 del, 17404 sub ] exp/tri2/decode_dev_beamformit_ref/wer_17_1.0
+
+# tri3
+%WER 72.81 [ 42869 / 58881, 3629 ins, 15998 del, 23242 sub ] exp/tri3/decode_dev_worn/wer_15_1.0
+%WER 91.73 [ 54013 / 58881, 3519 ins, 27098 del, 23396 sub ] exp/tri3/decode_dev_beamformit_ref/wer_17_1.0
+
+# nnet3 tdnn+chain
+%WER 47.91 [ 28212 / 58881, 2843 ins, 8957 del, 16412 sub ] exp/chain_train_worn_u100k_cleaned/tdnn1a_sp/decode_dev_worn/wer_9_0.0
+%WER 81.28 [ 47859 / 58881, 4210 ins, 27511 del, 16138 sub ] exp/chain_train_worn_u100k_cleaned/tdnn1a_sp/decode_dev_beamformit_ref/wer_9_0.5
+
+# result with the challenge submission format (July 9, 2018)
+# before the fix of speaker ID across arrays
+session S02 room DINING: #words 8288, #errors 6593, wer 79.54 %
+session S02 room KITCHEN: #words 12696, #errors 11096, wer 87.39 %
+session S02 room LIVING: #words 15460, #errors 12219, wer 79.03 %
+session S09 room DINING: #words 5766, #errors 4651, wer 80.66 %
+session S09 room KITCHEN: #words 8911, #errors 7277, wer 81.66 %
+session S09 room LIVING: #words 7760, #errors 6023, wer 77.61 %
+overall: #words 58881, #errors 47859, wer 81.28 %
+
+# result with the challenge submission format (July 9, 2018)
+# after the fix of speaker ID across arrays
+==== development set ====
+session S02 room DINING: #words 8288, #errors 6556, wer 79.10 %
+session S02 room KITCHEN: #words 12696, #errors 11096, wer 87.39 %
+session S02 room LIVING: #words 15460, #errors 12182, wer 78.79 %
+session S09 room DINING: #words 5766, #errors 4648, wer 80.61 %
+session S09 room KITCHEN: #words 8911, #errors 7277, wer 81.66 %
+session S09 room LIVING: #words 7760, #errors 6022, wer 77.60 %
+overall: #words 58881, #errors 47781, wer 81.14 %

egs/chime5/s5b/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="retry.pl queue.pl --mem 2G"
+export decode_cmd="queue.pl --mem 4G"
+

egs/chime5/s5b/conf/beamformit.cfg

@@ -0,0 +1,50 @@
+#BeamformIt sample configuration file for AMI data (http://groups.inf.ed.ac.uk/ami/download/)
+
+# scrolling size to compute the delays
+scroll_size = 250
+
+# cross correlation computation window size
+window_size = 500
+
+#amount of maximum points for the xcorrelation taken into account
+nbest_amount = 4
+
+#flag wether to apply an automatic noise thresholding 
+do_noise_threshold = 1
+
+#Percentage of frames with lower xcorr taken as noisy
+noise_percent = 10
+
+######## acoustic modelling parameters
+
+#transition probabilities weight for multichannel decoding
+trans_weight_multi = 25
+trans_weight_nbest = 25
+
+###
+
+#flag wether to print the feaures after setting them, or not
+print_features = 1
+
+#flag wether to use the bad frames in the sum process
+do_avoid_bad_frames = 1
+
+#flag to use the best channel (SNR) as a reference
+#defined from command line
+do_compute_reference = 1
+
+#flag wether to use a uem file or not(process all the file)
+do_use_uem_file = 0
+
+#flag wether to use an adaptative weights scheme or fixed weights
+do_adapt_weights = 1
+
+#flag wether to output the sph files or just run the system to create the auxiliary files
+do_write_sph_files = 1
+
+####directories where to store/retrieve info####
+#channels_file = ./cfg-files/channels
+
+#show needs to be passed as argument normally, here a default one is given just in case
+#show_id = Ttmp
+

egs/chime5/s5b/conf/mfcc.conf

@@ -0,0 +1,2 @@
+--use-energy=false
+--sample-frequency=16000

egs/chime5/s5b/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 
+--num-mel-bins=40
+--num-ceps=40
+--low-freq=40
+--high-freq=-400

egs/chime5/s5b/conf/online_cmvn.conf

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

egs/chime5/s5b/local/chain/run_tdnn.sh

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