hiroga-speech / README.md
Hiroaki Ogasawara
chore: label
fbe0878
metadata
license: apache-2.0

データセット作成手順

撮影

  1. QuickTime Playerを開く
  2. 画面から30cm程に近づく。距離が遠いと、LFROI(Lower-half Face ROI)作成時に顔全体が映り込んでしまうため
  3. 録画ボタンを押し、1秒待ってから発話。そうしないと初めが途切れる
  4. データ加工前に再生し、切れていないことを確認

データ加工

rawlfroiは次のとおり。なおLFROI_映像は音声を含まないので注意。

uvx --python 3.12 --from 'git+https://github.com/xhiroga/audio-visual-toolkit#subdirectory=packages/avt' crop-mouth --video-file raw/hiroga_001.mov --out-dir lfroi --height 300 --width 300
# or
uvx --python 3.12 --from 'git+https://github.com/xhiroga/audio-visual-toolkit#subdirectory=packages/avt' crop-mouth --video-dir raw --out-dir lfroi

raw.labは次のとおり。

mfa align raw japanese_mfa japanese_mfa aligned --output_format json --clean --num-jobs 8 --single_speaker
uvx --python 3.12 --from 'git+https://github.com/xhiroga/audio-visual-toolkit#subdirectory=packages/avt' mfa-aligned-to-label --json-dir aligned --out-dir labels