icefall/egs/spgispeech/ASR/prepare.sh

#!/usr/bin/env bash

# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python

set -eou pipefail

nj=20
stage=-1
stop_stage=100

# We assume dl_dir (download dir) contains the following
# directories and files. If not, they will be downloaded
# by this script automatically.
#
#  - $dl_dir/spgispeech
#      You can find train.csv, val.csv, train, and val in this directory, which belong
#      to the SPGISpeech dataset.
#
#  - $dl_dir/musan
#      This directory contains the following directories downloaded from
#       http://www.openslr.org/17/
#
#     - music
#     - noise
#     - speech
dl_dir=$PWD/download

. shared/parse_options.sh || exit 1

# vocab size for sentence piece models.
# It will generate data/lang_bpe_xxx,
# data/lang_bpe_yyy if the array contains xxx, yyy
vocab_sizes=(
  500
)

# All files generated by this script are saved in "data".
# You can safely remove "data" and rerun this script to regenerate it.
mkdir -p data

log() {
  # This function is from espnet
  local fname=${BASH_SOURCE[1]##*/}
  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
}

log "dl_dir: $dl_dir"

if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
  log "Stage 0: Download data"

  # If you have pre-downloaded it to /path/to/spgispeech,
  # you can create a symlink
  #
  #   ln -sfv /path/to/spgispeech $dl_dir/spgispeech
  #
  if [ ! -d $dl_dir/spgispeech/train.csv ]; then
    lhotse download spgispeech $dl_dir
  fi

  # If you have pre-downloaded it to /path/to/musan,
  # you can create a symlink
  #
  #   ln -sfv /path/to/musan $dl_dir/
  #
  if [ ! -d $dl_dir/musan ]; then
    lhotse download musan $dl_dir
  fi
fi

if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
  log "Stage 1: Prepare SPGISpeech manifest (may take ~1h)"
  # We assume that you have downloaded the SPGISpeech corpus
  # to $dl_dir/spgispeech. We perform text normalization for the transcripts.
  mkdir -p data/manifests
  lhotse prepare spgispeech -j $nj --normalize-text $dl_dir/spgispeech data/manifests
fi

if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
  log "Stage 2: Prepare musan manifest"
  # We assume that you have downloaded the musan corpus
  # to data/musan
  mkdir -p data/manifests
  lhotse prepare musan $dl_dir/musan data/manifests
  lhotse combine data/manifests/recordings_{music,speech,noise}.json data/manifests/recordings_musan.jsonl.gz
  lhotse cut simple -r data/manifests/recordings_musan.jsonl.gz data/manifests/cuts_musan_raw.jsonl.gz
fi

if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
  log "Stage 3: Split train into train and dev and create cut sets."
  python local/prepare_splits.py
fi

if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
  log "Stage 4: Compute fbank features for spgispeech dev and val"
  mkdir -p data/fbank
  python local/compute_fbank_spgispeech.py --test
fi

if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
  log "Stage 5: Compute fbank features for train"
  mkdir -p data/fbank
  python local/compute_fbank_spgispeech.py --train --num-splits 20

  log "Combine features from train splits (may take ~1h)"
  if [ ! -f data/manifests/cuts_train.jsonl.gz ]; then
    pieces=$(find data/manifests -name "cuts_train_[0-9]*.jsonl.gz")
    lhotse combine $pieces data/manifests/cuts_train.jsonl.gz
  fi
  gunzip -c data/manifests/cuts_train.jsonl.gz | shuf | gzip -c > data/manifests/cuts_train_shuf.jsonl.gz
fi

if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
  log "Stage 6: Compute fbank features for musan"
  mkdir -p data/fbank
  python local/compute_fbank_musan.py
fi

if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
  log "Stage 7: Dump transcripts for LM training"
  mkdir -p data/lm
  gunzip -c data/manifests/cuts_train_raw.jsonl.gz \
    | jq '.supervisions[0].text' \
    | sed 's:"::g' \
    > data/lm/transcript_words.txt
fi

if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
  log "Stage 8: Prepare BPE based lang"

  for vocab_size in ${vocab_sizes[@]}; do
    lang_dir=data/lang_bpe_${vocab_size}
    mkdir -p $lang_dir

    # Add special words to words.txt
    echo "<eps> 0" > $lang_dir/words.txt
    echo "!SIL 1" >> $lang_dir/words.txt
    echo "<UNK> 2" >> $lang_dir/words.txt

    # Add regular words to words.txt
    gunzip -c data/manifests/cuts_train_raw.jsonl.gz \
      | jq '.supervisions[0].text' \
      | sed 's:"::g' \
      | sed 's: :\n:g' \
      | sort \
      | uniq \
      | sed '/^$/d' \
      | awk '{print $0,NR+2}' \
      >> $lang_dir/words.txt

    # Add remaining special word symbols expected by LM scripts.
    num_words=$(cat $lang_dir/words.txt | wc -l)
    echo "<s> ${num_words}" >> $lang_dir/words.txt
    num_words=$(cat $lang_dir/words.txt | wc -l)
    echo "</s> ${num_words}" >> $lang_dir/words.txt
    num_words=$(cat $lang_dir/words.txt | wc -l)
    echo "#0 ${num_words}" >> $lang_dir/words.txt

    ./local/train_bpe_model.py \
      --lang-dir $lang_dir \
      --vocab-size $vocab_size \
      --transcript data/lm/transcript_words.txt

    if [ ! -f $lang_dir/L_disambig.pt ]; then
      ./local/prepare_lang_bpe.py --lang-dir $lang_dir
    fi
  done
fi

if [ $stage -le 9 ] && [ $stop_stage -ge 9 ]; then
  log "Stage 9: Train LM"
  lm_dir=data/lm

  if [ ! -f $lm_dir/G.arpa ]; then
    ./shared/make_kn_lm.py \
      -ngram-order 3 \
      -text $lm_dir/transcript_words.txt \
      -lm $lm_dir/G.arpa
  fi

  if [ ! -f $lm_dir/G_3_gram.fst.txt ]; then
    python3 -m kaldilm \
      --read-symbol-table="data/lang_phone/words.txt" \
      --disambig-symbol='#0' \
      --max-order=3 \
      $lm_dir/G.arpa > $lm_dir/G_3_gram.fst.txt
  fi
fi

if [ $stage -le 10 ] && [ $stop_stage -ge 10 ]; then
  log "Stage 10: Compile HLG"
  ./local/compile_hlg.py --lang-dir data/lang_phone

  for vocab_size in ${vocab_sizes[@]}; do
    lang_dir=data/lang_bpe_${vocab_size}
    ./local/compile_hlg.py --lang-dir $lang_dir
  done
fi