Datasets:

FBK-MT
/

fama-data

# Install sox (example for Ubuntu/Debian)
sudo apt-get install sox

# Install yt-dlp, more detailed instructions at: https://github.com/yt-dlp/yt-dlp/wiki/Installation
pip install -U "yt-dlp[default]"

Files Included

SplitAudioUsingSileroLog.pl - Perl script that processes audio using Silero logs
train_youtubecommons-en.ids - List of English file IDs to download
yt-commons-en.silero.json.gz - Compressed Silero log for English
train_youtubecommons-it.ids - List of Italian file IDs to download
yt-commons-it.silero.json.gz - Compressed Silero log for Italian

⚠️ Time markers in the final tsv files stored in this repository refer to the reduced audio files, not the original YouTube audio files.

Instructions

Follow the steps below to generate the audio segments starting from the logs available in this folder. If you are interested in replicating the logs, also follow the optional step for generating the logs.

Download Audio Files

Download in the folder ${DOWNLOAD_DIR} the audio files listed in ${VIDEO_IDS} using yt-dlp by running:

for id in `cat ${VIDEO_IDS}` ; do
    download_output=${DOWNLOAD_DIR}/${id}.wav
    if ! test -f  ${download_output} ; then
    echo "Saving audio track in ${download_output}"
    ${YT-DLP-PATH}/yt-dlp  \
        -r 500K \
        --cookies-from-browser firefox \
        --extract-audio \
        --audio-format wav \
        --postprocessor-args "-ar 16000 -ac 1" \
        -o ${download_output} \
        "https://www.youtube.com/watch?v=${id}"
    else
    echo "Skipping... ${download_output} already saved"
    fi
done

Where ${VIDEO_IDS} is train_youtubecommons-en.ids for English, and train_youtubecommons-it.ids for Italian.

Note: Some videos may no longer be available on YouTube, resulting in a subset of the original dataset. If an original audio file is missing due to download failure, it will be automatically skipped during processing.

(Optional) Generation of the Voice Activity Detection Logs

To reproduce the logs, install Silero, which serves to remove non-speech phenomena (silence, noise, music):

pip install silero-vad

Once installed, run the script speech_only.py present in this folder:

python ./speech_only.py \
    --folder ${DOWNLOAD_FOLDER} \
    --sfx ${WAV_SUFFIX} \
    --out_folder ${OUT_FOLDER} \
    --out_file   ${OUT_JSON_FILE}

The script processes the audio files in ${DOWNLOAD_FOLDER} with suffix ${WAV_SUFFIX} and stores the VAD-processed audios (in wav format at 16kHz) in ${OUT_FOLDER} along with the associated segmentation file (in json format) in ${OUT_JSON_FILE}.

Segmentation based on Voice Activity Detection Logs

Segment the audio downloaded in ${DOWNLOAD_DIR} using the logs ${LOG_FILE} and store them in ${AUDIO_SEGMENT_DIR} by running:

perl ./SplitAudioUsingSileroLog.pl ${LOG_FILE} ${DOWNLOAD_DIR} ${AUDIO_SEGMENT_DIR}

Where ${LOG_FILE} is yt-commons-en.silero.json.gz for English, and yt-commons-it.silero.json.gz for Italian.

Audio Segmentation using SHAS

To split the reduced audio into segments with a controlled duration, download and install SHAS following the official README in the ${SHAS_ROOT} folder.

To generate the final segmentation, run:

python ./segment-ytc.py --not_strict \
  -wavs ${PATH_TO_WAVS} \                      
  -ckpt ${CHECKPOINT_PATH} \ 
  -yaml ${OUTPUT_YAML}

Where ${PATH_TO_WAVS} is the path to the wav files obtained from Silero and stored in ${AUDIO_SEGMENT_DIR}, ${CHECKPOINT_PATH} is the path to the SHAS Multilingual model, ${OUTPUT_YAML} is the path where to store the final audio segmentation saved as yaml files and used for training.

Transcription and Translation

Transcription is done following the same processing used for the MOSEL dataset. Translation is done following the same processing of the other ASR datasets, described in the FAMA data card.

License and Citation

Please refer to the original main FAMA data card for licensing information and citation.