casting-call

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

I record my calls as a stereo .mov: my mic on the left channel, everyone else on the right, plus the Meet window on screen. This repo is everything that happens after the call ends. It started as "split the channels and run Whisper" and has grown into four 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. It is written to be read with zero prior knowledge, covers every command and constant, and has animated diagrams of the three core mechanisms.

The four capabilities

1. Transcribe (ripa). Split the stereo audio, run whisper-cpp on each channel, merge by timestamp into [h:mm:ss] [You] / [Caller] lines. Every remote participant lands on the right channel 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.

4. Measure delivery (ripvibe). A transcript flattens "that's valid" said warm and "that's valid" said flat into the same seven characters. The affect layer measures per utterance: pitch, energy, and pace from audio (via Praat), or smile intensity and nods from the speaker's video tile (via MediaPipe) when their audio is gone. Everything is z-scored against that speaker's own baseline, and the output is a list of moments worth re-listening to, not a mind reader.

And one safety net for the next recording: callcheck verifies both channels are actually capturing, either on an existing file (sampled scan, ~20s even on multi-GB recordings) or live at the start of a call. bin/callcheck_live.command is a Stream-Deck-ready front end: point a "System: Open" key at it and you get a 6-second capture plus a PASS/FAIL notification before the call gets going.

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)
ripvibe	analyze_affect.sh	prosody z-scores, or --faces smile/nod tracking
ript	transcribe_batch_stereo.sh	batch transcribe across dated folders
ripv / riprec	convert_video.sh / convert_recordings.sh	compress video / audio+video housekeeping
callcheck	callcheck.sh	verify both channels (file, --fast, --full, or --live)

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

callcheck --live                                   # before the call: is the rig capturing both sides?
# ...record the call with QuickTime + Loopback...
ripa ~/calls/2026-07-06                            # transcribe everything in the folder
rips recording.mov --region 120,1850,900,180 \
     --transcript transcripts/recording.txt        # put real names on the Caller lines
ripvibe recording.mov \
     --transcript transcripts/recording.txt        # which moments deserve a re-listen

And when callcheck would have failed but you did not run it:

ripcap recording.mov --region 120,1850,900,180     # words, from the captions
ripvibe recording.mov --faces --tile 150,95,1055,920   # delivery, from their face

Setup

brew install tesseract ffmpeg whisper-cpp
pip install --break-system-packages pillow numpy pytesseract pytest
# affect layer:
pip install --break-system-packages praat-parselmouth mediapipe

Whisper model (the batch scripts default to the smaller -q5_0 variant; the error message tells you which to fetch):

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'

Face landmark model (only for ripvibe --faces):

mkdir -p ~/.cache/casting-call
curl -sL -o ~/.cache/casting-call/face_landmarker.task \
  "https://storage.googleapis.com/mediapipe-models/face_landmarker/face_landmarker/float16/latest/face_landmarker.task"

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
  prosody.py       pitch/energy/pace features, within-speaker z-scores
  faces.py         smile blendshapes + nod detection
  affect.py        CLI orchestrator for prosody + faces
  tests/           pytest suite (fast, no fixtures on disk)
docs/user-guide.html   the real documentation; start there

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