casting-call

Turn call recordings into transcripts that know who said what, and how they said it.

I record my calls with OBS into a single .mkv: my mic on one track, the far side (Chrome/Meet) on a second, in-call markers on a third, plus the Meet window on screen. This repo is everything that happens after the call ends. It started as "split the tracks and run Whisper" and has grown into three capabilities that cover each way a recording can be useful (or broken).

New here or coming back cold? Read the user guide in a browser. Setting the rig up on a new machine is the setup guide. The user guide is written to be read with zero prior knowledge, covers every command and constant, and has animated diagrams of the three core mechanisms.

The three capabilities

1. Transcribe (ripa). Separate the audio tracks, run whisper-cpp on each, merge by timestamp into [h:mm:ss] [You] / [Caller] lines. Every remote participant lands on the far-side track as one undifferentiated Caller.

2. Attribute speakers (rips). Meet's on-screen captions already say who is speaking. Sample the video, OCR the caption band, fuzzy-match names against your local roster, build a speaker timeline, and rewrite Caller lines with real names. Includes an interactive review step for names it will not guess at.

3. Recover from a dead channel (ripcap). When the far-end audio never made it to disk (it happens; ask me how I know, twice), the captions burned into the video are the words. The stitcher OCRs every caption frame and merges the rolling, overlapping text into one attributed transcript.

Markers

Seven keys on a Stream Deck, each wired to a silent audio clip that lands on Track 3. Press one during a call and it shows up in the transcript as a [MARKER] line at that timestamp. Nobody on the call hears anything, and there is no syncing step, because all three tracks share one clock.

Key	Type	Press it when
Topic Switch	topic	the conversation moves to a new subject
Action For Me	action-me	you just picked up something to do
Action For Them	action-them	the far side owes you something
Important	important	it matters and fits nothing else
Question	question	you have a follow-up (say it into your muted mic right after)
Quote	quote	worth repeating verbatim
Video	video	something on screen worth a screenshot

Both action clips say the word "action", so the parser checks "for them" and "for me" before falling back to a bare action type. That fallback has no key. It only catches a press where Whisper ate the trailing word, and it surfaces as owner-unknown instead of landing on the wrong person's list (which is the failure you would never notice).

Wiring the OBS sources and the Stream Deck keys is in the setup guide, including a migration section if you already ran the old flag/action/follow set.

Commands

Alias	Script	Does
ripa	extract_audio_stereo.sh	extract + transcribe + auto-relabel if a timeline exists
rips	extract_speakers.sh	caption OCR → speaker timeline → relabel transcript
ripcap	stitch_captions.sh	captions → full transcript (dead-channel fallback)
ripv	convert_video.sh	re-encode to a smaller H.265 playable copy, audio mixed you=L / caller=R

The aliases live in my gutils shell config, not here. Every script works without them: bash bin/<script>.sh ... from the repo root.

Typical session

# ...record the call with OBS (multi-track: mic, far side, markers)...
ripa ~/calls/2026-07-06                            # transcribe everything in the folder
rips recording.mkv --region 120,1850,900,180 \
     --transcript recording/recording.txt          # put real names on the Caller lines

And when the far-side audio never made it to disk:

ripcap recording.mkv --region 120,1850,900,180     # words, from the captions

Setup

brew install tesseract ffmpeg whisper-cpp
pip install --break-system-packages pillow numpy pytesseract pytest

Whisper model (ripa defaults to ggml-large-v3.bin; the error message tells you which to fetch if it is missing):

mkdir -p ~/whisper-models
curl -o ~/whisper-models/ggml-large-v3.bin -L \
  'https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3.bin?download=true'

Roster (your real names stay local; the file is gitignored):

cp speakers_roster.example.json speakers_roster.json
# edit: set "self" to your name, list the people you call

Repo shape

bin/               shell entry points (thin; the logic lives in the package)
casting_call/      the Python package
  sample/locate/read/roster/timeline/transcript/review/coverage   speaker layer
  stitch.py        rolling-caption stitcher + CLI
  markers.py       Track 3 marker phrases folded into the transcript
  digest.py        marker presses grouped into a per-call digest (JSON)
  tests/           pytest suite (fast, no fixtures on disk)
docs/user-guide.html   the real documentation; start there
docs/setup-guide.html  building the OBS + Stream Deck rig on a new machine

Everything runs locally. No accounts, no uploads, and anything the tool cannot confidently attribute stays labeled Caller rather than being guessed.