Skip to content

Annotate

Audio#

The audio recordings should be sampled with 16kHz, in mono and stored in WAV files.

You can do this using wai.annotations. E.g., the following command converts all WAV files in INPUT_DIR and stores them in OUTPUT_DIR:

docker run --rm -u $(id -u):$(id -g) \
  -v `pwd`:/workspace \
  -t waikatoufdl/wai.annotations:0.8.0 \
  wai-annotations convert \
    from-audio-files-ac \
      -i "INPUT_DIR/*.wav" \
    convert-to-mono \
    resample-audio \
      -s 16000 \
    to-audio-files-ac \
      -o OUTPUT_DIR

Annotations#

The simplest file formats for speech annotations are probably:

  • Festvox
  • Coqui TTS

Festvox#

The annotations are stored in a text file, each row for a separated recording, using the following format:

( FILENAME "TRANSCRIPT" )
  • FILENAME - the name of the file without extension
  • TRANSCRIPT - the text associated with the recording

It is best to store the annotations file alongside the WAV files to avoid paths.

Coqui TTS#

In this format, the annotations are stored in a CSV file with the following columns:

  • wav_filename - the name of the file with extension
  • wav_filesize - the size of the WAV file in bytes
  • transcript - the text associated with the recording

It is best to store the annotations file alongside the WAV files to avoid paths.