FFmpeg script to convert VoxCeleb 2 dataset to wav files

[mogwai]

Having an explicit script to do this would help people downloading this dataset convert the files ready to be used in pyannote audio

[hbredin]

Or maybe find a way to avoid this conversion (though it would probably result in slower training because of on-the-fly m4a decoding).

[sangeetaghangam]

I used the following commands on bash, for some reason it did not let me use the file name substring so I ended up renaming the files

cmd from m4a to wav

find ./ -iname '*.m4a' -exec bash -c 'ffmpeg -i "{}" "{}".wav' ;

convert from m4a.wav to .wav

find . -name "*.m4a.wav" -exec rename 's/.m4a.wav$/.wav/' '{}' +

[mogwai]

I dug into this rabbit hole a while ago and found a nice script that we could adapt to do this:

#!/bin/bash

# In order to use this script, you must install parallel and
# ffmpeg

export SOURCE_DIR="/home/$USER/audiom4a"
export TARGET_DIR="/home/$USER/audiowav"

doone() {
    audioFile="$1"
    tmpVar="${audioFile%.*}.wav"
    wavFile="${tmpVar/$SOURCE_DIR/$TARGET_DIR}"
    wavFilePath=$(dirname "${audioFile}")
    mkdir -p "${wavFilePath}"
    if [ ! -f "$wavFile" ]; then # If the mp3 file doesn't exist already
        echo "Input: $audioFile"
        echo "Output: $wavFile"
        ffmpeg -i "$audioFile" "$wavFile" < /dev/null
    fi
}

export -f doone

find "${SOURCE_DIR}" -type f \( -iname "*.m4a" -or -iname "*.mp3" \) -print0 |
  parallel -0 doone