icefall/egs/ami/SURT/prepare.sh

#!/usr/bin/env bash

set -eou pipefail

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/ami
#      You can find audio and transcripts for AMI in this path.
#
#  - $dl_dir/icsi
#      You can find audio and transcripts for ICSI in this path.
#
#  - $dl_dir/rirs_noises
#      This directory contains the RIRS_NOISES corpus downloaded from https://openslr.org/28/.
#
dl_dir=$PWD/download

. shared/parse_options.sh || exit 1

# 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
vocab_size=500

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/amicorpus,
  # you can create a symlink
  #
  #   ln -sfv /path/to/amicorpus $dl_dir/amicorpus
  #
  if [ ! -d $dl_dir/amicorpus ]; then
    for mic in ihm ihm-mix sdm mdm8-bf; do
      lhotse download ami --mic $mic $dl_dir/amicorpus
    done
  fi

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

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

if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
  log "Stage 1: Prepare AMI manifests"
  # We assume that you have downloaded the AMI corpus
  # to $dl_dir/amicorpus. We perform text normalization for the transcripts.
  mkdir -p data/manifests
  for mic in ihm ihm-mix sdm mdm8-bf; do
    log "Preparing AMI manifest for $mic"
    lhotse prepare ami --mic $mic --max-words-per-segment 30 --merge-consecutive $dl_dir/amicorpus data/manifests/
  done
fi

if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
  log "Stage 2: Prepare ICSI manifests"
  # We assume that you have downloaded the ICSI corpus
  # to $dl_dir/icsi. We perform text normalization for the transcripts.
  mkdir -p data/manifests
  log "Preparing ICSI manifest"
  for mic in ihm ihm-mix sdm; do
    lhotse prepare icsi --mic $mic $dl_dir/icsi data/manifests/
  done
fi

if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
  log "Stage 3: Prepare RIRs"
  # We assume that you have downloaded the RIRS_NOISES corpus
  # to $dl_dir/rirs_noises
  lhotse prepare rir-noise -p real_rir -p iso_noise $dl_dir/rirs_noises data/manifests
fi

if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
  log "Stage 3: Extract features for AMI and ICSI recordings"
  python local/compute_fbank_ami.py
  python local/compute_fbank_icsi.py
fi

if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
  log "Stage 5: Create sources for simulating mixtures"
  # In the following script, we speed-perturb the IHM recordings and extract features.
  python local/compute_fbank_ihm.py
  lhotse combine data/manifests/ami-ihm_cuts_train.jsonl.gz \
    data/manifests/icsi-ihm_cuts_train.jsonl.gz - |\
    lhotse cut trim-to-alignments --type word --max-pause 0.5 - - |\
    lhotse filter 'duration<=12.0' - - |\
    shuf | gzip -c > data/manifests/ihm_cuts_train_trimmed.jsonl.gz
fi

if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
  log "Stage 6: Create training mixtures"
  lhotse workflows simulate-meetings \
    --method conversational \
    --same-spk-pause 0.5 \
    --diff-spk-pause 0.5 \
    --diff-spk-overlap 1.0 \
    --prob-diff-spk-overlap 0.8 \
    --num-meetings 200000 \
    --num-speakers-per-meeting 2,3 \
    --max-duration-per-speaker 15.0 \
    --max-utterances-per-speaker 3 \
    --seed 1234 \
    --num-jobs 2 \
    data/manifests/ihm_cuts_train_trimmed.jsonl.gz \
    data/manifests/ai-mix_cuts_clean.jsonl.gz

  python local/compute_fbank_aimix.py

  # Add source features to the manifest (will be used for masking loss)
  # This may take ~2 hours.
  python local/add_source_feats.py

  # Combine clean and reverb
  cat <(gunzip -c data/manifests/cuts_train_clean_sources.jsonl.gz) \
    <(gunzip -c data/manifests/cuts_train_reverb_sources.jsonl.gz) |\
    shuf | gzip -c > data/manifests/cuts_train_comb_sources.jsonl.gz
fi

if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then
  log "Stage 7: Create training mixtures from real sessions"
  python local/prepare_ami_train_cuts.py
  python local/prepare_icsi_train_cuts.py

  # Combine AMI and ICSI
  cat <(gunzip -c data/manifests/cuts_train_ami.jsonl.gz) \
    <(gunzip -c data/manifests/cuts_train_icsi.jsonl.gz) |\
    shuf | gzip -c > data/manifests/cuts_train_ami_icsi.jsonl.gz
fi

if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then
  log "Stage 8: Dump transcripts for BPE model training (using AMI and ICSI)."
  mkdir -p data/lm
  cat <(gunzip -c data/manifests/ami-sdm_supervisions_train.jsonl.gz | jq '.text' | sed 's:"::g') \
      <(gunzip -c data/manifests/icsi-sdm_supervisions_train.jsonl.gz | jq '.text' | sed 's:"::g') \
  > data/lm/transcript_words.txt
fi

if [ $stage -le 9 ] && [ $stop_stage -ge 9 ]; then
  log "Stage 9: Prepare BPE based lang (combining AMI and ICSI)"

  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
  cat data/lm/transcript_words.txt | grep -o -E '\w+' | sort -u | 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
fi