icefall/egs/emilia/TTS/prepare.sh

#!/usr/bin/env bash

set -eou pipefail

# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674
export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
# pip install lhotse s3tokenizer
stage=6
stop_stage=6

dl_dir=$PWD/download
dl_dir=/workspace_data/Emilia-Dataset/
prefix="emilia"
# zh, en, ja, ko, de, fr
lang_set=("de" "en" "zh" "ja" "ko" "fr")
lang_set=("de" "en" "zh" "ja" "fr")
. 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]}) $*"
}

if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
  log "dl_dir: $dl_dir"
  log "Stage 0: Download data"
  #huggingface-cli login
  # huggingface-cli download --repo-type dataset --local-dir $dl_dir Wenetspeech4TTS/WenetSpeech4TTS

  # Extract the downloaded data:
  for lang in "${lang_set[@]}"; do
    lang_upper=$(echo "${lang}" | tr '[:lower:]' '[:upper:]')
    folder=$dl_dir/raw/${lang_upper}
    for file in $folder/*.tar.gz; do
      echo "Processing ${file}"
      # e.g. $dl_dir/raw/DE/*tar.gz untar first, DE is the language code in upper case
      tar -xzvf $file -C $folder
    done
  done
fi

if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
  log "Stage 1: Prepare emilia manifest"
  # We assume that you have downloaded the Emilia corpus
  # to $dl_dir/emilia
  mkdir -p data/manifests
  for lang in "${lang_set[@]}"; do
    echo "Processing ${lang}"
    if [ ! -e data/manifests/.emilia.${lang}.done ]; then
      lhotse prepare emilia $dl_dir data/manifests --num-jobs 30 --lang "${lang}"
      touch data/manifests/.emilia.${lang}.done
    fi
  done
fi


if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
  log "Stage 2: Generate fbank (used by ./f5-tts)"
  mkdir -p data/fbank
  for lang in "${lang_set[@]}"; do
    echo "Processing ${lang}"
    if [ ! -e data/fbank/.emilia.${lang}.done ]; then
      ./local/compute_mel_feat.py --dataset-parts $lang --split 100 --prefix ${prefix}
      touch data/fbank/.emilia.${lang}.done
    fi
  done
fi

if [ $stage -le 16 ] && [ $stop_stage -ge 16 ]; then
  log "Stage 6: Split the ${prefix} cuts into train, valid and test sets (used by ./f5-tts)"
  if [ ! -f data/fbank/${prefix}_cuts_${subset}.jsonl.gz ]; then
    echo "Combining ${prefix} cuts"
    pieces=$(find data/fbank/ -name "${prefix}_cuts_${subset}.*.jsonl.gz")
    lhotse combine $pieces data/fbank/${prefix}_cuts_${subset}.jsonl.gz
  fi
  if [ ! -e data/fbank/.${prefix}_split.done ]; then
    echo "Splitting ${prefix} cuts into train, valid and test sets"

    lhotse subset --last 800 \
      data/fbank/${prefix}_cuts_${subset}.jsonl.gz \
      data/fbank/${prefix}_cuts_validtest.jsonl.gz
    lhotse subset --first 400 \
      data/fbank/${prefix}_cuts_validtest.jsonl.gz \
      data/fbank/${prefix}_cuts_valid.jsonl.gz
    lhotse subset --last 400 \
      data/fbank/${prefix}_cuts_validtest.jsonl.gz \
      data/fbank/${prefix}_cuts_test.jsonl.gz

    rm data/fbank/${prefix}_cuts_validtest.jsonl.gz

    n=$(( $(gunzip -c data/fbank/${prefix}_cuts_${subset}.jsonl.gz | wc -l) - 800 ))
    lhotse subset --first $n  \
      data/fbank/${prefix}_cuts_${subset}.jsonl.gz \
      data/fbank/${prefix}_cuts_train.jsonl.gz
      touch data/fbank/.${prefix}_split.done
  fi
fi

# zcat test.jsonl.gz | jq -r '.recording.id + " " + .recording.sources[0].source' > wav.scp
if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
  log "Stage 4: Extract cosyvoice2 FSQ token (used by ./f5-tts semantic token experiment)"
  data_dir=$dl_dir/raw/ZH
  # for all jsonl files in data_dir
  for jsonl_file in $data_dir/*.jsonl; do
    # get the file basename
    jsonl_file_basename=$(basename $jsonl_file)
    echo "Processing $jsonl_file"
    output_dir="./cosy_v2_tokens_ZH/${jsonl_file_basename%.jsonl}"
    echo "output_dir: $output_dir"
    # skip if the output_dir exists
    if [ -e $output_dir ]; then
      echo "Output directory $output_dir already exists, skipping"
      continue
    fi
    mkdir -p $output_dir
    torchrun --nproc_per_node=8 --nnodes=1 \
        --rdzv_id=2024 --rdzv_backend="c10d" --rdzv_endpoint="localhost:0" \
        local/extract_cosyvoice2_token.py --data_dir $data_dir \
                    --jsonl_file $jsonl_file_basename \
                    --device "cuda" \
                    --output_dir $output_dir \
                    --batch_size 32 \
                    --num_workers 2 \
                    --model "speech_tokenizer_v2_25hz"  # or "speech_tokenizer_v1_25hz
  done
fi

if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
  log "Stage 5: Extract cosyvoice2 FSQ token (used by ./f5-tts semantic token experiment)"
  for lang in "${lang_set[@]}"; do
    lang_upper=$(echo "${lang}" | tr '[:lower:]' '[:upper:]')
    data_dir=$dl_dir/raw/${lang_upper}
    # for all jsonl files in data_dir
    for jsonl_file in $data_dir/*.jsonl; do
    # get the file basename
      jsonl_file_basename=$(basename $jsonl_file)
      echo "Processing $jsonl_file"
      output_dir="./cosy_v2_tokens_${lang_upper}/${jsonl_file_basename%.jsonl}"
      echo "output_dir: $output_dir"
      # skip if the output_dir exists
      if [ -e $output_dir ]; then
        echo "Output directory $output_dir already exists, skipping"
        continue
      fi
      mkdir -p $output_dir
      torchrun --nproc_per_node=8 --nnodes=1 \
          --rdzv_id=2024 --rdzv_backend="c10d" --rdzv_endpoint="localhost:0" \
          local/extract_cosyvoice2_token.py --data_dir $data_dir \
                      --jsonl_file $jsonl_file_basename \
                      --device "cuda" \
                      --output_dir $output_dir \
                      --batch_size 32 \
                      --num_workers 2 \
                      --model "speech_tokenizer_v2_25hz"  # or "speech_tokenizer_v1_25hz
    done
  done
fi

if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
# cat EN_B00008.tar.gz.* > EN_B00008.tar.gz
  for lang in "${lang_set[@]}"; do
    lang_upper=$(echo "${lang}" | tr '[:lower:]' '[:upper:]')
    cosy_token_dir="./cosy_v2_tokens_${lang_upper}"
    for dir in $cosy_token_dir/*; do
      echo "Processing $dir"
      # get the file basename
      dir_basename=$(basename $dir)
      echo "dir_basename: $dir_basename"
      cat $dir/part* > $dir/${dir_basename}.jsonl
    done
    cat $cosy_token_dir/${lang_upper}*/*.jsonl > $cosy_token_dir/cosy_v2_tokens_${lang_upper}.jsonl
  done
fi