avenox-roughcut

Rough cut — transcript-driven (silence + flubs)

Two kinds of cut: silence (dead air, mechanical → auto-editor) and flubs/retakes (a restarted sentence, semantic → transcript + agent judgment). The director approves the flub list before anything is cut.

Pipeline

Job dir (LOCAL — never inside a synced/cloud folder): $STUDIO_JOBS/<job>/{raw,cut,transcript,frames}

Set STUDIO_JOBS to wherever you keep heavy media, e.g. export STUDIO_JOBS=~/video/projects. Keeping media out of a synced folder matters: cloud sync will thrash on multi-GB intermediates.

1. Transcribe (local mlx-whisper — Apple Silicon)

cd "$STUDIO_JOBS/<job>"
python3 -c "
import os, certifi; os.environ['SSL_CERT_FILE']=certifi.where(); os.environ['REQUESTS_CA_BUNDLE']=certifi.where()
import mlx_whisper, json
r=mlx_whisper.transcribe('<RAW>', path_or_hf_repo='mlx-community/whisper-large-v3-turbo', language='<LANG>', word_timestamps=False)
segs=[{'i':i,'start':round(s['start'],2),'end':round(s['end'],2),'text':s['text'].strip()} for i,s in enumerate(r['segments'])]
json.dump({'text':r['text'].strip(),'segments':segs}, open('transcript/raw_timed.json','w'), ensure_ascii=False, indent=1)
"

~48s for 14 min of audio on an M-series Mac. Local is the default — it is faster and cheaper than any API round trip at this length. Note that most LLM-routing proxies have no whisper endpoint; if you must go remote, use a dedicated speech API.

2. Detect flubs (read transcript, propose to the director)

Scan raw_timed.json for:

• repeated sentence-starts (the same opening said twice)
• cut-off restarts (a half sentence, then the full take)
• self-corrections ("we need X" → "instead of X, …")
• hanging filler words right before a gap

Present as a table (mm:ss + text). The director approves before cutting. This step stays human-gated — an agent cutting semantic content unreviewed will eventually remove a real point.

3. Cut — flubs (ffmpeg) THEN silence (auto-editor)