This repository contains a Node.js script to combine multiple transcript files into a single file. The script sorts the transcript lines based on timestamps and prefixes each line with the character's name.
I built this to help make it easier when summarizing D&D campaign session audio. By putting all the players' transcripts together, it helps a large language model (LLM) understand the flow of events better and creates a more accurate summary.
- Clone this repository
- Install the dependencies using
npm ci
You can run the script using the following command:
npm run tool:combine -- --output ./tmp/combined-transcripts.txt \
--dedupe "consecutive" \
--skip-filter "/\[AUDIO OUT\]|\[BLANK_AUDIO\]/" \
--timestamped false \
--player-name "DM" --role "Dungeon Master" --character-name "DM" --character-description "DM for the campaign" --transcript "./tmp/dm-transcript.vtt" \
--player-name "Player 1" --role "Player" --character-name "Player 1" --character-description "Fighter" --transcript "./tmp/player-1-transcript.vtt" \
--player-name "Player 2" --role "Player" --character-name "Player 2" --character-description "Ranger" --transcript "./tmp/player-2-transcript.vtt" \
--player-name "Player 3" --role "Player" --character-name "Player 3" --character-description "Wizard" --transcript "./tmp/player-3-transcript.vtt"The --dedupe option specifies the deduplication strategy to be used when processing VTT files. It accepts the following values:
"false": No deduplication is performed. All content is included as-is."consecutive": Removes consecutive duplicate content. This is useful for cleaning up transcripts where the same message might be repeated multiple times in a row."unique": Ensures only the first occurrence of each unique message is included. This is useful for eliminating repeated content that may occur due to AI hallucinations.
The --skip-filter option allows you to specify a series of content strings or regex patterns to filter out messages from the parsed VTT files. If a message matches any of the specified filters, it will be excluded from the output. This is useful for removing unwanted content such as [AUDIO OUT] or [BLANK_AUDIO].
You can pass multiple filters as strings:
--skip-filter "[AUDIO OUT]" "[BLANK_AUDIO]"Or as a single regex pattern:
--skip-filter "/\[AUDIO OUT\]|\[BLANK_AUDIO\]/"The --timestamped option is a boolean flag that controls whether timestamps are included in the output. When enabled (default), timestamps are included. When disabled, timestamps are omitted from the output.
The --chunks option specifies the number of parts to split the combined transcript into. When set to 1 (default), no splitting occurs. When set to a value greater than 1, the combined transcript is divided into the specified number of parts, with each part saved as a separate file.
To generate transcripts for D&D live-play sessions, I use a combination of things.
- Craig: A Discord bot used to record the sessions.
- FFmpeg: A tool to convert audio files from ogg to wav format.
- Install via Homebrew:
brew install ffmpeg
- Install via Homebrew:
- whisper.cpp: A tool for transcribing audio files.
Use Craig via Discord to record the sessions, then download them in the ogg format.
Convert the ogg files to wav format using FFmpeg.
for f in *.ogg; do ffmpeg -i "$f" -ar 16000 -ac 1 "${f%.ogg}.wav"; doneUse whisper.cpp to transcribe the wav files into VTT format.
for f in *.wav; do ~/Projects/whisper.cpp/main \
-m ~/Projects/whisper.cpp/models/ggml-large-v3.bin \
-f "$f" \
--output-srt \
--print-colors \
--print-progress \
--output-file "${f%.wav}" \
--beam-size 5 \
--entropy-thold 2.8 \
--max-context 64 \
--prompt "This is a live-play Dungeons & Dragons 5e session."
done- Automating D&D Notetaking with AI: An article that inspired this process.