Interview, tightened

Videos:

• final.mp4: 9:16 for Reels · repo copy at 720x1280 (the master is 1080x1920) · 30 fps · 72.3 s · 16.5 MB · -14.0 LUFS · bold-pop captions burned in
• final-16x9.mp4: 16:9 · 1920x1080 · 30 fps · 72.3 s · 17.6 MB · -14.0 LUFS · phrase-level clean captions burned in, with final-16x9.srt as a sidecar

These are the round-3 versions, after two critic reviews. See Review round below.

The request#

"Cut the ums and long pauses out of this interview clip, add captions, and make a vertical 9:16 version for Reels plus a 16:9 version"

Mode: quick. No questions were asked. The opening line stated the assumptions: keep every sentence in order, remove fillers, shorten pauses over 0.5 s to 0.3 s, use face-tracked 9:16 with bold-pop captions for Reels and 16:9 with clean captions, apply a light grade, master at -14 LUFS, and add no music because the piece is led by speech.

The "user's footage" is an 81 s excerpt of a real public-domain NASA interview. It was found with showtime assets media search (NASA source, video). It is a sit-down interview with astronaut Christina Koch, recorded in 2023 for the Artemis II crew announcement, taken from source 1:01-2:22 of NASA/JSC item jsc2023m000095. See Sources and license below.

What it demonstrates#

• Editing by transcript. The edit is driven by a word-level transcript (Whisper large-v3-turbo, 262 words). The cut list comes from edit cut. The hand edits to the EDL were the punch-ins and reframe splits, the grade strength, the caption options, in round 2 a 0.17 s pad before "and readiness" and a 0.22 s tail trim, and in round 3 a fixed crop on one punch-in range.
• Fillers and pauses. The filler pass ran and found 0 ums. Parakeet, the engine that keeps disfluencies, confirmed it. NASA's interview was cleanly delivered, so nothing was invented or forced. The pause pass removed the lead-in, the tail and five pauses of 0.6-1.3 s. The clip went from 81.0 s to 72.3 s (-11 %). A Parakeet re-transcription of the final still has all 262 words, so no word was clipped at a cut. Made with showtime 0.1.0. showtime 0.2.0's default transcription (Parakeet v3 plus a filler scan) finds two "uh" in this clip that 0.1.0's tools missed ("more raw uh technical", "great, and uh I hope"); this example was not re-made.
• Hidden jump cuts. Alternate ranges punch in to 1.2x, so each of the 5 real cuts reads as a change of framing. The last 35 s (one unbroken answer) also changes framing 4 times without a cut: the range is split into contiguous, frame-aligned source ranges that alternate between 1.0x and 1.2x, with each join placed in a gap between words. These joins carry no audio fade, so the room tone runs on unbroken (see Review round). See project/review/scenes.jpg.
• Face-tracked 9:16 reframe. YuNet face tracking with a dead zone keeps her centred without drift. On one punch-in (0:37.1-0:42.1) the dead zone held the crop on where she started, which left her at about 60 % of the width, so that range uses a fixed crop on her face instead ("fit": "cover", "focus": {"x": 0.503}).
• Caption placement is a judgment call, and it was checked in the frames. In the first draft, bold-pop's default lower third sat exactly on the flight-suit patch and name tag. The captions moved to position: middle, on the collar just under the chin, and in round 2 grew to about an 84 px cap height (size 0.135, chars 16 so the longest line still fits). In round 3 the bold-pop chunks come from boldpop_captions.py (showtime's own bold-pop pass, plus a rule that no chunk is shorter than 0.20 s). The 16:9 version uses sentence-case phrase captions in the lower 8 %, and every caption change lands on the cut, not 3 frames after it.
• Light grade. teal-orange at strength 0.3. grade --compare stills showed that --grade auto lifted the low-key backdrop and flattened her face, so auto was rejected.
• Loudness. Mastered once on the program to -14 LUFS with a -1 dBTP ceiling.

Commands, in order#

showtime is skills/showtime/bin/showtime, run from the working folder. <job> is showtime-out/nasa-interview-edit-20260926-140703. The helper scripts run with showtime's Python (PYTHONPATH=skills/showtime/lib ~/.showtime/venv/bin/python).

showtime doctor --quick

# find real footage: a NASA interview with a person on camera
showtime assets media search "astronaut talks about" --type video --source nasa --limit 12 --preview sheet1.jpg
showtime assets media search "astronaut interview"   --type video --source nasa --limit 12 --preview sheet2.jpg
showtime assets media search "NASA Astronaut Candidate" --type video --source nasa --json   # + more queries
#   (durations/sizes checked from each item's images-assets.nasa.gov metadata.json; see friction notes)

showtime job init nasa-interview-edit --platform reels \
  --goal "Cut the ums and long pauses out of this interview clip, add captions, and make a vertical 9:16 version for Reels plus a 16:9 version"
showtime assets media fetch "nasa:jsc2023m000095_Christina_Koch_Artemis_II_Crew_Announcement_Resource_Reel-social" \
  --quality large --max-mb 400 -o raw/
showtime footage probe raw/jsc2023m000095_...-social.mp4                   # 1920x1080, 30 fps, 8:02, stereo AAC
showtime footage scenes raw/jsc2023m000095_...-social.mp4 --every 10 --job <job>   # interview runs 1:01-2:25

# the "user's clip": a one-range EDL (project/raw/excerpt/edl.json), no grade, no loudness
showtime edit render raw/excerpt/edl.json -o raw/koch-interview.mp4

showtime job note <job> --stage understand --verified "..." --assumed "..." --next transcribe
showtime transcribe raw/koch-interview.mp4 --edit-dir <job>/edit --speakers 1 \
  --prompt "Christina Koch, Artemis II, Kennedy Space Center, mission specialist"
showtime pack <job>/edit                                                     # read takes_packed.md
showtime transcribe raw/koch-interview.mp4 --edit-dir <job>/work/parakeet --model parakeet   # filler cross-check: 0

# look choice
showtime footage grade raw/koch-interview.mp4 --analyze
showtime footage grade raw/koch-interview.mp4 --auto --compare --at 20 -o gradetest/auto.png
showtime footage grade raw/koch-interview.mp4 --look teal-orange --strength 0.35 --compare --at 20 -o gradetest/teal.png
showtime footage grade raw/koch-interview.mp4 --look clean-punch --strength 0.4 --compare --at 20 -o gradetest/punch.png
showtime footage grade raw/koch-interview.mp4 --preset subtle --compare --at 20 -o gradetest/subtle.png

# EDL, then draft -> look -> fix (3 passes)
showtime job note <job> --stage plan --verified "..." --assumed "..." --next "edit cut"
showtime edit cut <job>/edit/transcripts/koch-interview.json --max-pause 0.5 \
  --aspect 9:16 --captions bold-pop --grade teal-orange -o <job>/edit/edl.json
#   hand edit: "grade": {"lut": "teal-orange", "strength": 0.3}, "loudness": {"lufs": -14, "tp": -1}
showtime edit check <job>
showtime edit render <job> --preview && showtime edit view <job>             # pass 1: captions on the chest patch
#   hand edit: "zoom" on ranges 2, 4, 6
showtime edit render <job> --preview --overwrite && showtime edit view <job> # pass 2: patch still under captions
#   hand edit: "captions": {"style": "bold-pop", "position": "middle"}
showtime edit render <job> --preview --overwrite                             # pass 3: clean

# round-1 finals, qa and review pack (see "Review round" for what changed)
showtime edit render <job>/edit/edl-16x9.json -o <job>/final-16x9.mp4 --overwrite
showtime edit render <job>/edit/edl.json -o <job>/final.mp4
showtime qa <job> && showtime review-pack <job>

# round 2: apply the critic's findings
#   EDLs: range 3 start 16.37 -> 16.20, range 6 split into 5 contiguous 1.0x/1.2x ranges, tail 77.88 -> 77.66,
#   zoom 1.1 -> 1.2; 9:16 captions {"size": {"portrait": 0.135}, "chars": 16}; 16:9 {"captions": false,
#   "subtitles": "captions-16x9.ass"}
python project/rechunk_captions.py <job>/edit/edl-16x9.json --srt final-16x9.srt \
  --ass <job>/edit/captions-16x9.ass --style clean --position bottom        # 25 phrase cues, shortest 1.12 s
showtime edit check <job>/edit/edl.json
showtime edit render <job>/edit/edl.json -o <job>/final-2.mp4
showtime edit render <job>/edit/edl-16x9.json -o <job>/final-16x9-2.mp4
python project/rechunk_captions.py <job>/edit/edl.json --srt <job>/final-2.srt   # replaces the render's SRT
showtime deliver thumb <job>/final-2.mp4 --at 50 --size 540x960 --fit scale -o round2/after/v_50.jpg   # before/after stills
showtime qa <job>/final-2.mp4 --platform reels                               # WARN: upscale, caption_fast
showtime qa <job>/final-16x9-2.mp4 --platform youtube                        # WARN: caption_fast
showtime transcribe <job>/final-2.mp4 --edit-dir <job>/work/qa-transcript2 --model parakeet   # 262 words
showtime deliver poster <job>/final-2.mp4 --at 25.5
showtime review-pack <job>/final-2.mp4 --platform reels                      # round 2
showtime review-pack <job>/final-16x9-2.mp4 --platform youtube -o <job>/review/16x9
showtime job note <job> --stage feedback --verified "round 2 applied: ..."

# round 3: second critic pass (no blockers)
#   showtime: edit render now skips the 30 ms edge fades between contiguous ranges of one source
#   EDLs: range 6 {"fit": "cover", "focus": {"x": 0.503, "y": 0.25}}; 9:16 {"captions": false,
#   "subtitles": "captions-916.ass"}
python project/boldpop_captions.py <job>/edit/edl.json <job>/edit/captions-916.ass \
  --size 0.135 --chars 16 --position middle                                 # "OUR SKILLS", "THAT THAT"
python project/rechunk_captions.py <job>/edit/edl-16x9.json --srt <job>/final-16x9-3.srt \
  --ass <job>/edit/captions-16x9.ass --style clean --position bottom        # 25 cues, starts on cuts
python project/rechunk_captions.py <job>/edit/edl.json --srt <job>/final-3.srt   # 35 cues
showtime edit render <job>/edit/edl.json -o <job>/final-3.mp4               # 5 of 10 segments re-rendered
showtime edit render <job>/edit/edl-16x9.json -o <job>/final-16x9-3.mp4
showtime deliver thumb <job>/final-3.mp4 --at 39.567 --size 540x960 --fit scale -o round3/after/v_39.567.jpg
showtime qa <job>/final-3.mp4 --platform reels                               # WARN: caption_fast, upscale
showtime qa <job>/final-16x9-3.mp4 --platform youtube                        # WARN: caption_fast
showtime transcribe <job>/final-3.mp4 --edit-dir <job>/work/qa-transcript3 --model parakeet   # 262 words
showtime review-pack <job>/final-3.mp4 --platform reels --force-round        # cut sheet only; no critic

# deliver
showtime deliver exports <job>/final-3.mp4 --targets reels                   # upload master, 48.7 MB
showtime deliver exports <job>/final-16x9-3.mp4 --targets youtube            # upload master, 47.7 MB
python project/export_repo.py <job>/final-3.mp4 final.mp4 reels 720x1280     # 16.5 MB repo copy
python project/export_repo.py <job>/final-16x9-3.mp4 final-16x9.mp4 youtube  # 17.6 MB repo copy
showtime qa final.mp4 --platform reels && showtime qa final-16x9.mp4 --platform youtube   # WARN, WARN
showtime assets credits credtmp --all -o credits.txt   # credtmp/ holds only the source's .license.json
showtime job note <job> --stage deliver --verified "..." --next done

export_repo.py calls st.deliver.exports.export_one with the Reels/YouTube target set to CRF 22 and a 1.85 Mbps video cap (160 kbps audio), keeping showtime's loudnorm, BT.709 tags and faststart. It is not a raw ffmpeg call. It exists because deliver exports has no size-budget option. The 9:16 repo copy is 720x1280: at this bitrate a 1080x1920 frame smeared skin, and the face-tracked crop holds only about 608x1080 real pixels anyway. The full-quality masters from deliver exports (1080x1920 and 1920x1080) are what you would upload.

rechunk_captions.py regroups the same output-timeline words into phrase cues with a small dynamic program: at most 42 characters per line on 16:9 (32 on 9:16) and 2 lines per cue, breaks preferred after sentences, clauses and pauses, no line or cue ending on a function word ("the", "of", "to", "when" ...) or splitting a name when another break fits, and each cue held at least 1.0 s. A cue that would start up to 250 ms after a cut or reframe starts on it instead. It writes the SRT, and the ASS through showtime's own captions.to_ass in the clean style. It exists because showtime's grouping split this fast speaker into 80-200 ms one-word cues (see Review round).

boldpop_captions.py runs showtime's bold-pop grouping unchanged, then merges any chunk shorter than 0.20 s with a neighbour ("OUR" joins "SKILLS"; "THAT" takes the next "THAT"). It writes the ASS the 9:16 EDL burns through "subtitles".

Timings#

Measured on a 6-core Intel Mac (x64) whose CPU was shared with about 3 other render jobs.

QA summary (round 3)#

• 9:16 master (<job>/final-3.mp4, 1080x1920, 72.27 s), verdict WARN: 0 fail, 2 warn. File h264 High, yuv420p, 30 fps, faststart, BT.709. Loudness -14.0 LUFS, true peak -1.4 dBTP. No silent gaps, black or frozen stretches.
  • WARN upscale: the source is enlarged 1.78x, or 2.13x on the 1.2x punch-in ranges. That is unavoidable when a 1080p source fills a 1080x1920 frame, and NASA's original is also 1080p. Round 2 raised the punch-in from 1.1x (1.96x) to 1.2x on purpose, so the framing changes read as deliberate.
  • WARN caption_fast on the SRT sidecar: phrase cues run over 20 characters/s in her fastest passages (18 of 35 cues). She averages about 19 characters/s across the whole clip, so verbatim phrase captions can't all stay under 20. Round 3 made the 9:16 sidecar favour whole phrases over reading speed, which added cues over 20 (10 of 28 in round 2). The round-1 SRT passed this check only by splitting such passages into 1-2 word flashes, which the rule exempts.
• 16:9 master (<job>/final-16x9-3.mp4), verdict WARN: 0 fail, 1 warn (caption_fast, same cause, 10 of 25 cues). Enlargement: 1.2x on the punch-in ranges (under qa's 1.5x limit), 1.0x elsewhere. Captions: 25 cues, shortest 1.12 s, none under 1 s (round 1: 85 cues, 8 under 200 ms).
• Shipped copies, verdict WARN for both. final.mp4 against reels: resolution (the 720x1280 repo copy, see above) and caption_fast. final-16x9.mp4 against youtube: caption_fast. qa measures both at -14.0 LUFS / -1.5 dBTP.
• Words: a Parakeet re-transcription of final-3.mp4 has all 262 words.
• Review packs: <job>/review/round-1/ (the first critic's input), <job>/review/round-2/ and <job>/review/16x9/round-1/ (the second critic's input), and <job>/review/round-3/ (built with --force-round for the shipped cut sheet; no critic reviewed round 3).
• Not checked: listening. Audio was judged from the waveform, the loudness graph, RMS at each join and a re-transcription.

Review round#

Round 1 was self-reviewed and judged "ship". A separate critic then reviewed the round-1 pack plus the 16:9 and the SRTs. Its verdict was ship after fixes: no blockers, 4 should-fix items and 6 polish items. Each item was checked against the frames or files before anything changed. All of them were applied in one re-render per aspect. The log is in <job>/work/feedback.md.

What the fixes cost:

• The upscale went from 1.96x to 2.13x on the punch-in ranges.
• The honest caption_fast WARN replaced a PASS that the one-word flashes had earned by exploiting the rule's exemption.
• The shipped 9:16 copy now carries a resolution WARN.
• Two bold-pop chunks ("OUR" at 53.3 s and "THAT" at 61.2 s) lasted 0.15-0.16 s, shorter than round 1's 0.20 s minimum (fixed in round 3).
• Each reframe-only join had a 30 ms audio fade at both edges, which dipped the room tone for about 20 ms (fixed in round 3).

Round 3#

A second critic reviewed the round-2 pack, the 16:9 pack and the SRTs. Its verdict was again ship after fixes: no blockers, 1 should-fix and 7 polish items. Each was checked against the files or frames first; all were applied. The before/after stills are in the job's round3/ folder.

Round 3 was not reviewed again: review.md allows two critic rounds.

What went wrong on the way#

• The first 16:9 final burned captions mid-frame. edit render --captions clean replaces the style but keeps the EDL's position: middle from the 9:16 plan. The fix was a separate edl-16x9.json, and it cost one 245 s render.
• The first two footage candidates were 79-minute satellite "live shots" (3+ GB). Search results show no duration or size, so each candidate's metadata was checked by hand.
• The round-1 self-review missed the caption flashes. review-pack packs only the latest final, qa's caption check passed them, and the 16:9 was never looked at frame by frame. The critic found them.
• The first split of the last range at 47.45 s overlapped the previous range by 10 ms, because each range's end is snapped to whole frames from its own start. The joins were re-set to frame-aligned times before rendering.

Files#

• final.mp4, final.srt: the Reels version (720x1280 repo copy) and its phrase-level SRT (optional for Reels; the captions are burned in).
• final-16x9.mp4, final-16x9.srt: the 16:9 version and its sidecar for YouTube/LinkedIn.
• poster.jpg: the frame at 0:25.5 ("ALWAYS DREAMED"), 1080x1920.
• share.txt: Reels and YouTube copy, plus the credit line.
• credits.txt: the source credit and NASA usage notes.
• project/edit/edl.json: the 9:16 EDL, and project/edit/captions-916.ass, the bold-pop file it burns. project/edit/edl-16x9.json is the 16:9 EDL, and project/edit/captions-16x9.ass is the caption file it burns.
• project/edit/transcripts/koch-interview.json: the word-level transcript.
• project/edit/takes_packed.md: the packed read of the transcript.
• project/raw/excerpt/edl.json: how the 81 s clip was cut from NASA's reel.
• project/raw/*.license.json: the license record from the fetch (sha256, URLs).
• project/rechunk_captions.py: the phrase-level caption chunker.
• project/boldpop_captions.py: the bold-pop pass with the 0.20 s chunk floor.
• project/export_repo.py: the repo-size export.
• project/review/scenes.jpg: the round-3 cut sheet (every scene middle, and each cut at -0.1 s, midpoint and +0.2 s).

The source media is not committed. To rebuild, re-fetch it with the assets media fetch line above, render the excerpt EDL to project/raw/koch-interview.mp4, run boldpop_captions.py on edl.json and rechunk_captions.py on edl-16x9.json, then showtime edit render each EDL. The transcript cache is keyed by content, so the transcribe step re-uses the committed transcript only when the rendered clip is identical; otherwise re-run transcribe.

Sources and license#

• Footage: "Christina Koch Artemis II Crew Announcement Resource Reel", NASA/JSC, 2023 (NASA Image and Video Library item jsc2023m000095_Christina_Koch_Artemis_II_Crew_Announcement_Resource_Reel-social), https://images.nasa.gov/details/jsc2023m000095_Christina_Koch_Artemis_II_Crew_Announcement_Resource_Reel-social. NASA's item description says the reel "includes interview footage and b-roll recorded in 2023". Only the interview portion (1:01-2:22) is used. She speaks about the mission as upcoming, so the post copy says "2023 interview".
• License: NASA-produced media, public domain (not subject to US copyright), under NASA's media usage guidelines: https://www.nasa.gov/nasa-brand-center/images-and-media/. No attribution is legally required; it is given anyway.
• No endorsement implied. NASA did not make, review or endorse this edit or showtime. The NASA insignia appears only as the patch on her flight suit in the original footage, so it is visible in every frame, including the poster. The edit adds no NASA logo, title or card. The words are Christina Koch's own: nothing was re-ordered or removed except pauses and silence.
• Fonts: Anton (bold-pop) and Inter (clean), both SIL OFL, bundled by showtime.

The project files on GitHub