icefall/egs/himia/wuw/prepare.sh

#!/usr/bin/env bash

set -eou pipefail

stage=0
stop_stage=6

# HI_MIA and aishell dataset are used in this experiment.
# musan dataset is used for data augmentation.
#
# For aishell dataset downloading and preparation,
# refer to icefall/egs/aishell/ASR/prepare.sh.
#
# For HI_MIA and HI_MIA_CW dataset,
# we assume dl_dir (download dir) contains the following
# directories and files. If not, they will be downloaded
# by this script automatically.
# Then these files will be extracted to $dl_dir/HiMia/
#
#  - $dl_dir/train.tar.gz
#      Himia training dataset.
#      From https://www.openslr.org/85
#
#  - $dl_dir/dev.tar.gz
#      Himia Devlopment dataset.
#      From https://www.openslr.org/85
#
#  - $dl_dir/test_v2.tar.gz
#      Himia test dataset.
#      From https://www.openslr.org/85
#
#  - $dl_dir/data.tgz
#      Himia confusion words(HI_MIA_CW) test dataset.
#      From https://www.openslr.org/120

#  - $dl_dir/resource.tgz
#      Transcripts of (HI_MIA_CW) test dataset.
#      From https://www.openslr.org/120

dl_dir=$PWD/download
train_set_channel=_7_01
enable_speed_perturb=False

. 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

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 HI_MIA and HI_MIA_CW dataset to /path/to/himia/,
  # you can create a symlink
  #
  #   ln -sfv /path/to/himia $dl_dir/
  #
  if [ ! -f $dl_dir/train.tar.gz ]; then
    lhotse download himia $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

  # If you have pre-downloaded it to /path/to/aishell,
  # you can create a symlink
  #
  #   ln -sfv /path/to/aishell $dl_dir/aishell
  #
  # The directory structure is
  # aishell/
  # |-- data_aishell
  # |   |-- transcript
  # |   `-- wav
  # `-- resource_aishell
  #     |-- lexicon.txt
  #     `-- speaker.info

  if [ ! -d $dl_dir/aishell/data_aishell/wav/train ]; then
    lhotse download aishell $dl_dir
  fi
fi

if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
  log "Stage 1: Prepare HI_MIA and HI_MIA_CW manifest"
  mkdir -p data/manifests
  if [ ! -e data/manifests/.himia.done ]; then
    lhotse prepare himia $dl_dir/HiMia data/manifests
    touch data/manifests/.himia.done
  fi
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
  if [ ! -e data/manifests/.musan.done ]; then
    lhotse prepare musan $dl_dir/musan data/manifests
    touch data/manifests/.musan.done
  fi
fi

if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
  log "Stage 3: Prepare aishell manifest"
  # We assume that you have downloaded the aishell corpus
  # to $dl_dir/aishell
  if [ ! -f data/manifests/.aishell_manifests.done ]; then
    mkdir -p data/manifests
    lhotse prepare aishell $dl_dir/aishell data/manifests
    touch data/manifests/.aishell_manifests.done
  fi
fi

if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
  log "Stage 4: Compute fbank for aishell"
  if [ ! -f data/fbank/.aishell.done ]; then
    mkdir -p data/fbank
    ./local/compute_fbank_aishell.py \
      --enable-speed-perturb=${enable_speed_perturb}
    touch data/fbank/.aishell.done
  fi
fi

if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
  log "Stage 5: Compute fbank for musan"
  mkdir -p data/fbank
  if [ ! -e data/fbank/.musan.done ]; then
    ./local/compute_fbank_musan.py
    touch data/fbank/.musan.done
  fi
fi

if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
  log "Stage 6: Compute fbank for HI_MIA and HI_MIA_CW dataset"
  # Format of train_set_channel is "micropohone position"_"channel"
  # Microphone 1 to 6 is an array with 16 channels.
  # Microphone 8 only has a single channel.
  # So valid examples of train_set_channel could be:
  # 1_01, ..., 1_16
  # 2_01, ..., 2_16
  # ...
  # 6_01, ..., 6_16
  # 7_01
  train_set_channel="_7_01"
  for subset in train dev test; do
    for file_type in recordings supervisions; do
      src=data/manifests/himia_${file_type}_${subset}.jsonl.gz
      dst=data/manifests/himia_${file_type}_${subset}${train_set_channel}.jsonl.gz
      cat <(gunzip -c ${src}) | \
        grep ${train_set_channel} | \
        gzip -c  > ${dst}
    done
  done

  mkdir -p data/fbank
  if [ ! -e data/fbank/.himia.done ]; then
    ./local/compute_fbank_himia.py \
      --train-set-channel=${train_set_channel} \
      --enable-speed-perturb=${enable_speed_perturb}
    touch data/fbank/.himia.done
  fi

  train_file=data/fbank/cuts_train_himia${train_set_channel}-aishell-shuf.jsonl.gz
  if [ ! -f ${train_file} ]; then
    # SingleCutSampler is preferred for this experiment
    # rather than DynamicBucketingSampler.
    # Since negative audios(Aishell) tends to be longer than positive ones(HiMia).
    # if DynamicBucketingSample is used, a batch may contain either all negative sample
    # or positive sample.
    # So `shuf` the training dataset here and use SingleCutSampler to load data.
    cat <(gunzip -c data/fbank/aishell_cuts_train.jsonl.gz) \
      <(gunzip -c data/fbank/cuts_train${train_set_channel}.jsonl.gz) | \
      grep -v _sp | \
      shuf |shuf | gzip -c > ${train_file}
  fi

fi