Barberotheca

Group project for the Semantic Digital Libraries 2025/26 course.

A digital library of Alessandro Barbero's lectures in audio and text form.

Process

Manual search

We searched on YouTube videos of Alessandro Barbero's lectures and compiled the barbero.csv file with basic metadata of each chosen source.

Audio download

We downloaded the audio-only version of each video using yt-dlp into the audio folder as M4A files.

# extract all youtube URLs into a batch file
tail -n +2 metadata/barbero.csv | cut -d',' -f1 > .yt-dlp/barbero.lst

# download m4a audio files with yt-dlp
yt-dlp \
  --format bestaudio[ext=m4a] \
  --extract-audio --audio-format m4a \
  --sleep-interval 30 --limit-rate 5M \
  --batch-file .yt-dlp/barbero.lst --download-archive .yt-dlp/barbero.log \
  --output "audio/barbero-%(extractor)s-%(id)s.%(ext)s"

We decided to use M4A format with AAC compression, as it provides better audio fidelity at smaller file sizes, which is beneficial for accurate transcription with Whisper, compared to MP3 which introduces more compression artifacts that could hurt transcription accuracy.

Semantic renaming

We compiled the barbero.csv with the reasoned semantic filenames, adding the column semantic_filename and filling it manually.
Then we mass-renamed all files in the audio folder to their semantic version using the script semantic_rename.py.

python scripts/semantic_rename.py

Whisper transcription

We transcribed each audio file using OpenAI Whisper in Italian language using the turbo model.

for file in audio/*.m4a; do
  whisper "$file" \
    --model turbo \
    --language it \
    --task transcribe \
    --output_format all \
    --output_dir transcripts/
done

We then did a quick quality check of the transcripts files. If for any reason the turbo model failed in transcribing some of the audio, we retried manually using the small model.

Keywords/entities extraction

We extracted keywords and named entities from each transcript txt file using the script keywords_extract.py which leverages SpaCy NLP model for italian (it_core_news_lg), counting all terms frequencies and selecting them based on their part-of-speech tags. We then saved the results as a JSON file keyword-entities.json containing the top 50 keywords and 30 entities for each filename.

python scripts/keywords_extract.py

Manual compilation of keywords

We choose manually the list of reasoned keywords for each lecture and we added them to barbero.csv file in the columns keywords and entities.

Audio compression

We compressed the audio files to reduce their size for easier storage and handling after the higher fidelity version, used for transcription, was no longer needed.
We chose to compress to AAC format with 48kbps bitrate, mono channel, and 22050Hz sample rate, which provides a good balance between audio quality and file size for spoken word content like lectures.

for file in audio/*.m4a; do
  ffmpeg -i "$file" \
    -c:a aac -b:a 48k \
    -ac 1 -ar 22050 \
    "compressed/$(basename "$f")"
done

CSV to JSON conversion

We converted the final barbero.csv file to JSON format using the script csv2json_convert.py which handles specific columns as integers or arrays.

Website development

Work in progress...

Disclaimer

This repository and all files contained within are used solely for educational purposes as part of a university project. No copyright infringement is intended. All media, texts, and materials remain the property of their respective owners and are included or referenced here only for academic, non-commercial use.

Team members

• Tommaso Barbato
• Martina Uccheddu