WIP [not-for-merge]: run aishell with latest recipe in Kaldi

egs/aishell/s10/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.
+--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) 

egs/aishell/s10/conf/online_cmvn.conf

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

egs/aishell/s10/local/nnet3/run_ivector_common.sh

@@ -0,0 +1,181 @@
+#!/bin/bash
+
+set -e -o pipefail
+
+
+# This script is called from local/nnet3/run_tdnn.sh and local/chain/run_tdnn.sh (and may eventually
+# be called by more scripts).  It contains the common feature preparation and iVector-related parts
+# of the script.  See those scripts for examples of usage.
+
+
+stage=0
+nj=30
+
+train_set=train_cleaned   # you might set this to e.g. train.
+gmm=tri3_cleaned          # This specifies a GMM-dir from the features of the type you're training the system on;
+                          # it should contain alignments for 'train_set'.
+online_cmvn_iextractor=false
+
+num_threads_ubm=8
+nnet3_affix=_cleaned     # affix for exp/nnet3 directory to put iVector stuff in, so it
+                         # becomes exp/nnet3_cleaned or whatever.
+
+. ./cmd.sh
+. ./path.sh
+. utils/parse_options.sh
+
+
+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
+  if [ ! -f $f ]; then
+    echo "$0: expected file $f to exist"
+    exit 1
+  fi
+done
+
+
+# lowres features, alignments
+if [ -f data/${train_set}_sp/feats.scp ] && [ $stage -le 2 ]; then
+  echo "$0: data/${train_set}_sp/feats.scp already exists.  Refusing to overwrite the features "
+  echo " to avoid wasting time.  Please remove the file and continue if you really mean this."
+  exit 1;
+fi
+
+if [ $stage -le 1 ]; then
+  echo "$0: preparing directory for low-resolution speed-perturbed data (for alignment)"
+  utils/data/perturb_data_dir_speed_3way.sh \
+    data/${train_set} data/${train_set}_sp
+
+  for datadir in ${train_set}_sp dev test; do
+    utils/copy_data_dir.sh data/$datadir data/${datadir}_hires
+  done
+fi
+
+if [ $stage -le 2 ]; then
+  echo "$0: making MFCC features for low-resolution speed-perturbed data"
+  steps/make_mfcc.sh --nj $nj \
+    --cmd "$train_cmd" data/${train_set}_sp
+  steps/compute_cmvn_stats.sh data/${train_set}_sp
+  echo "$0: fixing input data-dir to remove nonexistent features, in case some "
+  echo ".. speed-perturbed segments were too short."
+  utils/fix_data_dir.sh data/${train_set}_sp
+fi
+
+if [ $stage -le 3 ]; then
+  if [ -f $ali_dir/ali.1.gz ]; then
+    echo "$0: alignments in $ali_dir appear to already exist.  Please either remove them "
+    echo " ... or use a later --stage option."
+    exit 1
+  fi
+  echo "$0: aligning with the perturbed low-resolution data"
+  steps/align_fmllr.sh --nj $nj --cmd "$train_cmd" \
+         data/${train_set}_sp data/lang $gmm_dir $ali_dir
+fi
+
+
+if [ $stage -le 5 ] && [ -f data/${train_set}_sp_hires/feats.scp ]; then
+  echo "$0: data/${train_set}_sp_hires/feats.scp already exists."
+  echo " ... Please either remove it, or rerun this script with stage > 2."
+  exit 1
+fi
+
+if [ $stage -le 5 ]; then
+  echo "$0: creating high-resolution MFCC features"
+
+  # this shows how you can split across multiple file-systems.  we'll split the
+  # MFCC dir across multiple locations.  You might want to be careful here, if you
+  # have multiple copies of Kaldi checked out and run the same recipe, not to let
+  # them overwrite each other.
+  mfccdir=data/${train_set}_sp_hires/data
+  if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $mfccdir/storage ]; then
+    utils/create_split_dir.pl /export/b0{5,6,7,8}/$USER/kaldi-data/mfcc/tedlium-$(date +'%m_%d_%H_%M')/s5/$mfccdir/storage $mfccdir/storage
+  fi
+
+  # do volume-perturbation on the training data prior to extracting hires
+  # features; this helps make trained nnets more invariant to test data volume.
+  utils/data/perturb_data_dir_volume.sh data/${train_set}_sp_hires
+
+  for datadir in ${train_set}_sp dev test; do
+    steps/make_mfcc.sh --nj $nj --mfcc-config conf/mfcc_hires.conf \
+      --cmd "$train_cmd" data/${datadir}_hires
+    steps/compute_cmvn_stats.sh data/${datadir}_hires
+    utils/fix_data_dir.sh data/${datadir}_hires
+  done
+fi
+
+if [ $stage -le 6 ]; then
+  echo "$0: computing a subset of data to train the diagonal UBM."
+
+  mkdir -p exp/nnet3${nnet3_affix}/diag_ubm
+  temp_data_root=exp/nnet3${nnet3_affix}/diag_ubm
+
+  # train a diagonal UBM using a subset of about a quarter of the data
+  num_utts_total=$(wc -l <data/${train_set}_sp_hires/utt2spk)
+  num_utts=$[$num_utts_total/4]
+  utils/data/subset_data_dir.sh data/${train_set}_sp_hires \
+    $num_utts ${temp_data_root}/${train_set}_sp_hires_subset
+
+  echo "$0: computing a PCA transform from the hires data."
+  steps/online/nnet2/get_pca_transform.sh --cmd "$train_cmd" \
+    --splice-opts "--left-context=3 --right-context=3" \
+    --max-utts 10000 --subsample 2 \
+    ${temp_data_root}/${train_set}_sp_hires_subset \
+    exp/nnet3${nnet3_affix}/pca_transform
+
+  echo "$0: training the diagonal UBM."
+  # Use 512 Gaussians in the UBM.
+  steps/online/nnet2/train_diag_ubm.sh --cmd "$train_cmd" --nj 30 \
+    --num-frames 700000 \
+    --num-threads $num_threads_ubm \
+    ${temp_data_root}/${train_set}_sp_hires_subset 512 \
+    exp/nnet3${nnet3_affix}/pca_transform exp/nnet3${nnet3_affix}/diag_ubm
+fi
+
+if [ $stage -le 7 ]; then
+  # Train the iVector extractor. µUse all of the speed-perturbed data since iVector extractors
+  # can be sensitive to the amount of data. The script defaults to an iVector dimension of 100.
+  echo "$0: training the iVector extractor"
+  steps/online/nnet2/train_ivector_extractor.sh --cmd "$train_cmd" --nj 15 \
+    --num-threads 4 --num-processes 2 \
+    --online-cmvn-iextractor $online_cmvn_iextractor \
+    data/${train_set}_sp_hires exp/nnet3${nnet3_affix}/diag_ubm \
+    exp/nnet3${nnet3_affix}/extractor || exit 1;
+fi
+
+if [ $stage -le 8 ]; then
+  # note, we don't encode the 'max2' in the name of the ivectordir even though
+  # that's the data we extract the ivectors from, as it's still going to be
+  # valid for the non-'max2' data, the utterance list is the same.
+  ivectordir=exp/nnet3${nnet3_affix}/ivectors_${train_set}_sp_hires
+  if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $ivectordir/storage ]; then
+    utils/create_split_dir.pl /export/b0{5,6,7,8}/$USER/kaldi-data/ivectors/tedlium-$(date +'%m_%d_%H_%M')/s5/$ivectordir/storage $ivectordir/storage
+  fi
+  # We now extract iVectors on the speed-perturbed training data .  With
+  # --utts-per-spk-max 2, the script pairs the utterances into twos, and treats
+  # each of these pairs as one speaker; this gives more diversity in iVectors..
+  # Note that these are extracted 'online' (they vary within the utterance).
+
+  # Having a larger number of speakers is helpful for generalization, and to
+  # handle per-utterance decoding well (the iVector starts at zero at the beginning
+  # of each pseudo-speaker).
+  temp_data_root=${ivectordir}
+  utils/data/modify_speaker_info.sh --utts-per-spk-max 2 \
+    data/${train_set}_sp_hires ${temp_data_root}/${train_set}_sp_hires_max2
+
+  steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj $nj \
+    ${temp_data_root}/${train_set}_sp_hires_max2 \
+    exp/nnet3${nnet3_affix}/extractor $ivectordir
+
+  # Also extract iVectors for the test data, but in this case we don't need the speed
+  # perturbation (sp) or small-segment concatenation (comb).
+  for data in dev test; do
+    steps/online/nnet2/extract_ivectors_online.sh --cmd "$train_cmd" --nj "$nj" \
+      data/${data}_hires exp/nnet3${nnet3_affix}/extractor \
+      exp/nnet3${nnet3_affix}/ivectors_${data}_hires
+  done
+fi
+
+
+exit 0;

egs/aishell/s10/local/run_cleanup_segmentation.sh

@@ -0,0 +1,68 @@
+#!/bin/bash
+
+# Copyright 2016  Vimal Manohar
+#           2016  Johns Hopkins University (author: Daniel Povey)
+# Apache 2.0
+
+# This script demonstrates how to re-segment training data selecting only the
+# "good" audio that matches the transcripts.
+# The basic idea is to decode with an existing in-domain acoustic model, and a
+# biased language model built from the reference, and then work out the
+# segmentation from a ctm like file.
+
+# For nnet3 and chain results after cleanup, see the scripts in
+# local/nnet3/run_tdnn.sh and local/chain/run_tdnn.sh
+
+# GMM Results for speaker-independent (SI) and speaker adaptive training (SAT) systems on dev and test sets
+# [will add these later].
+
+set -e
+set -o pipefail
+set -u
+
+stage=0
+cleanup_stage=0
+data=data/train
+cleanup_affix=cleaned
+srcdir=exp/tri3
+nj=100
+decode_nj=16
+decode_num_threads=4
+
+. ./path.sh
+. ./cmd.sh
+. utils/parse_options.sh
+
+cleaned_data=${data}_${cleanup_affix}
+
+dir=${srcdir}_${cleanup_affix}_work
+cleaned_dir=${srcdir}_${cleanup_affix}
+
+if [ $stage -le 1 ]; then
+  # This does the actual data cleanup.
+  steps/cleanup/clean_and_segment_data.sh --stage $cleanup_stage --nj $nj --cmd "$train_cmd" \
+    $data data/lang $srcdir $dir $cleaned_data
+fi
+
+if [ $stage -le 2 ]; then
+  steps/align_fmllr.sh --nj $nj --cmd "$train_cmd" \
+    $cleaned_data data/lang $srcdir ${srcdir}_ali_${cleanup_affix}
+fi
+
+if [ $stage -le 3 ]; then
+  steps/train_sat.sh --cmd "$train_cmd" \
+    5000 100000 $cleaned_data data/lang ${srcdir}_ali_${cleanup_affix} ${cleaned_dir}
+fi
+
+if [ $stage -le 4 ]; then
+  # Test with the models trained on cleaned-up data.
+  utils/mkgraph.sh data/lang ${cleaned_dir} ${cleaned_dir}/graph
+
+  for dset in dev test; do
+    steps/decode_fmllr.sh --nj $decode_nj --num-threads $decode_num_threads \
+       --cmd "$decode_cmd"  --num-threads 4 \
+       ${cleaned_dir}/graph data/${dset} ${cleaned_dir}/decode_${dset}
+    steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" data/lang data/lang_rescore \
+       data/${dset} ${cleaned_dir}/decode_${dset} ${cleaned_dir}/decode_${dset}_rescore
+  done
+fi