Meeting transcription methodology
How to turn a BHG site walkthrough recording into notes, a transcript, stills, and glossary updates. First used on 28 August 2026. Follow this for later meetings; don’t skip the stills pass — audio alone mishears names and can’t see “this page”. Don’t publish until the sensitive-talk discussion has happened.
Spellings live in glossary. Extend that file as you correct; don’t invent a second glossary per meeting.
What you are producing
One folder per meeting: meetings/YYYY-MM-DD/.
| File | Job |
|---|---|
meeting-notes.md | The reliable account: decisions, walkthrough, actions. Write from audio first, then rewrite from stills. |
transcript.md | Timestamped diarized transcript. Evidence, not the cleaned story. Light ASR fixes only. |
screenshot-timestamps.json | Inclusive grab list {id, t, cue} for extraction. |
screenshot-timestamps.md | Same list, human-readable. |
stills/ | JPEGs NNN_HHMMSS.jpg. Working files; prefer not to commit. |
phone-stills/ | JPEGs taken on the phone during the meeting (books, cards, print-outs), not extracted from video. Prefer not to commit. |
stills-catalog.md | What each cluster of frames actually shows (and what the camera missed). Include phone stills here. |
extract_stills.py | Optional local copy; canonical script is <code>extract_stills.py</code> in this folder. Multiple clips need a clip field in the JSON (see 4 Sep). |
Also: a row in the meetings index, and any new names/IA in the project glossary. After notes exist and the sensitive-talk discussion has happened, rebuild BHG Site Working (python scripts/build.py in bhgsite-working) so the meeting is on the workshop site.
Do not put in git: the video, extracted wav/mp3 chunks, API keys, or (unless asked) the JPEG stills.
1. Recording
- Phone video is fine. Audio is primary; the camera is a pointing device.
- Say at the start that you’re recording so it can be transcribed later.
- Point at the monitor when someone says “this / there / that map / this list”. You will still miss stretches (desk, clothing, glare). That’s expected — stills tell you which timestamps were useful.
- Pixel files are often 1920×1080 with a display-matrix rotation (−90). ffmpeg autorotate is the right orientation for a portrait phone filming a landscape monitor. Don’t add an extra
transposeunless a test frame is sideways. - Keep the original file wherever it landed (
Downloads, etc.). Don’t copy 9 GB intodocs. - Phone stills (covers, cards, print-outs) are part of the evidence. Copy the JPEGs into
phone-stills/if that helps the catalog; don’t skip them because they aren’t in the video. - Sequential Pixel clips are separate clocks. Line them up by filename/UTC; don’t smash them into one 00:00 timeline if there is a gap. Label transcript timestamps
A/B(or similar).
Probe before transcribing:
ffprobe -v error -select_streams v:0 -show_entries stream=width,height,duration -show_entries stream_side_data=rotation -of json VIDEO.mp4
Note duration, rotation, and (if present) GPS / creation time for the notes header.
2. Audio → first transcript
Keys and models
- Native OpenAI STT is optional. This machine often has no
OPENAI_API_KEY. gpt-4o-transcribe-diarizeis not on OpenRouter — don’t wait for it.- Use Deepgram Nova-3 via OpenRouter:
POST https://openrouter.ai/api/v1/audio/transcriptions
model deepgram/nova-3
response_format=verbose_json
provider options: diarize, punctuate, utterances, smart_format.
- OpenRouter key from House of Ur deploy secrets. Never print it. Never commit it.
Chunking
Long files fail or stall. Split to ~8-minute 16 kHz mono wav, then 64 kbps mp3 chunks.
ffmpeg -i VIDEO.mp4 -ac 1 -ar 16000 audio.wav
# then cut into ~480s mp3 pieces
Work in %TEMP%\bhg-meeting-YYYYMMDD\ (or similar). Combined JSON can stay there; only markdown lands in docs.
Speaker IDs
Deepgram labels reset every chunk. Speaker 3:1 is not the same person as Speaker 0:1.
- In
transcript.md, keepSpeaker N:M(chunk N, speaker M). - Infer typical voices from content, not from IDs, and say so at the top of the transcript.
- Do not rewrite the whole transcript as named speakers. Guessing will be wrong after the first cut.
- In
meeting-notes.md, use real names (Emlyn / Barbara / Bob, etc.) from what was said.
Transcript conventions
- Timestamp every utterance:
[HH:MM:SS–HH:MM:SS] Speaker N:M. - First pass: obvious place-name fixes only (Burra, not Barra/Borrow/Borough) so the file is readable.
- Leave fillers, false starts, and uncertain words. The notes file is where you interpret.
- Header must say: model, chunking, speaker-ID warning, link to notes + glossary.
3. First notes (audio only)
Write meeting-notes.md before stills:
- Header: date, place, attendees, subject, recording filename, STT source.
- Purpose, decisions, overall reaction.
- Walkthrough in the order they actually went (not a tidy IA).
- Cutover must-haves vs later.
- Action table (who / what).
This pass will be wrong on exact page titles, fake group names, and “this list”. That’s fine. You need the spine so the timestamp list has somewhere to hang.
4. Screenshot timestamps (inclusive)
Read the transcript and list every moment the screen might add information. If in doubt, include it.
Include when someone:
- navigates (go to, click, back, previous, search)
- points (
this,there,that map,that box,hot link,scroll) - names a page, heading, booklet, or person on the page
- describes a bug you can’t hear (formatting, 404, casino hijack, duplicate rows)
Skip only if the span is clearly not about the screen (long policy talk with no click). Even then, one still at the start of the span is cheap.
Times are on the video clock, same as the transcript. Grab a couple of seconds into the span — they click, then point.
Write both:
{"id": "001", "t": "00:00:08", "cue": "Recording starts; maps / settlement by period"}
and a markdown table. Dense is correct: the 28 Aug pass used 185 grabs on a 73-minute tape.
5. Extract stills
Canonical script: <code>extract_stills.py</code>.
python meetings/extract_stills.py --video "C:\path\to\VIDEO.mp4" --meeting "C:\path\to\docs\projects\bhg-site\meetings\YYYY-MM-DD"
It reads screenshot-timestamps.json, seeks with ffmpeg (-ss before -i), autorotates, scales to 900px wide, JPEG q=3.
Sanity-check rotation on one known screen moment before the full run (four variants: default, -noautorotate, transpose=1, transpose=2). Keep the variant where webpage text is horizontal.
JPEG file size is a proxy:
| Size (this pipeline) | Likely content |
|---|---|
| ≳ 80 KB | Readable monitor |
| ~30–80 KB | Partial screen / glare / angle |
| ≲ 30 KB | Desk, paper, clothing, covered lens |
Don’t describe every dark frame. Catalog the large ones; sample the small ones to confirm they’re useless.
6. Look at the stills (this is the correction pass)
For each useful still, transcribe what is on the page:
- URL / breadcrumb / heading
- Visible nav (left menu items, tiles, “In this section”)
- Exact phrases they were arguing about (Inc., 46 plaques, “Download files available for purchase”, lowercase “indigenous”, “Burra Library”)
- Names on screen (Rosemary Ros, Mary Warnes, Elizabeth Doidge, Cyndi’s List)
- Layout bugs (doubled headings, duplicate table rows)
Write stills-catalog.md in two lists:
- On the monitor — still id, time, what’s there.
- Camera off — don’t pretend you saw Tindale / Features / Mongolata if the frame is a jumper.
When the camera missed, check the rebuild HTML (bhgsite2026/source/...) for the page they named. Creek Street “Vanstone, Hinch” was recovered that way on 28 Aug.
If a still contradicts the transcript (“Cherry and Kidd” vs GURNEY and KNAGGS on screen), keep both: they may have scrolled, or named a nearby heading. Then grep the site — 28 Aug had both headings, and the formatting bug was duplicate HTML blocks.
7. Correct the transcript
Rules:
- Glossary wins for established spellings (Burra, Ngadjuri, Ian Auhl, Mary Warnes, Rosemary Ros, Cyndi’s List, Yorke Peninsula).
- The still / HTML wins over ASR when the page is in frame (
Inc., plaque count, “Burra Family History Group”, book titles). - Don’t “clean up” speech. Only fix tokens that are clearly misheard names, places, or titles.
- Record in the transcript header which fixes you applied.
- Never remap
Speaker N:Mto names in the body.
Typical Deepgram traps on this project (fix on sight):
| ASR often says | Prefer |
|---|---|
| Barra, Borrow, Borough, Barrow | Burra |
| Warrens | Warnes (Mary Warnes / CWA) |
| Rosemary Ross | Rosemary Ros |
| Cindy’s List | Cyndi’s List |
| York Peninsula | Yorke Peninsula |
| ink | Inc. |
| paperwork (People hub) | People |
| Vanstown Hinch | Vanstone, Hinch |
| Bard Community School | Burra Community School |
| hIan Auhl | Ian Auhl |
Don’t “fix” a word into a glossary term unless the still or HTML supports it.
8. Improve the notes
Second pass on meeting-notes.md, driven by the catalog:
- Replace guessed page titles with on-screen titles and URLs.
- Quote the exact broken copy (46 plaques, shop prices, “from Burra Library”).
- Separate must-fix HTML bugs (duplicate surname sections, dead “Download files”) from content opinions.
- Mark what you didn’t see (Bob’s Nature and Heritage Walk map, Mongolata page).
- Link
stills-catalog.mdand the glossary from the Screen / Files section.
The notes remain the document you act on. The transcript stays the raw tape.
9. Sensitive and internal-only talk
The workshop site publishes notes and the transcript. That is a wider room than the people who were in Burra. After the notes and transcript exist, stop and discuss with Emlyn before publishing (rebuild of BHG Site Working, and any commit of those files). Do not skip this because the rest of the pipeline is green.
What to look for (from the 4 September pass; extend the list when a later tape shows a new class):
- Gossip or speculation about people or groups currently in town who are not a website decision.
- Funder talk: amounts, rates, “equivalent of Eric,” casino/sponsorship colour, unpaid specialist work as a public page.
- Local politics, constitution/incorporation disputes, and named current-town sketches that were not asked to go on the site.
- Business cards and contact lines photographed with no spoken decision (the IREN card is the type).
- 21 Sep additions: a named volunteer used as a funding-pitch example; Indigenous interview / who may visit a site on private land; other companies’ quiet-money precedents; ten-year ask figures (same class as funder amounts).
What to keep:
- Website decisions, walkthrough, and actions.
- Contributor names the group asked to record for People / methodology pages (Jodie’s list on 4 Sep).
- Who arrived, and ordinary meeting logistics.
How to leave it out:
- Keep the full notes and transcript in
docs(meetings/YYYY-MM-DD/). - Put the working-site copy in
published/meeting-notes.mdandpublished/transcript.mdin that same folder. Build prefers those files when they exist. - Notes (published): omit the span. Do not write that anything was redacted, withheld, or politically sensitive.
- Transcript (published): stub the same spans with
[omitted]if a cut would otherwise look like a gap in the clock; do not summarise what was omitted. - Quiet tickets (BHG-017, BHG-018) are reminders for later, not news items and not extra Features.
If unsure, leave it in the docs draft and ask. Fail closed: do not publish the meeting until that conversation has happened.
10. Glossary
After every meeting, add anything new to glossary:
- People (attendees and names on pages)
- Organisations and places
- Nav / URL patterns
- Publication titles
- External-link labels (Useful links)
- Recurring copy pitfalls
Prefer one row per term. Don’t duplicate the meeting notes. The glossary is for the next transcript, not a second set of minutes.
Companion for Country names: the Ngadjuri language note in planning docs.
11. Index
Add a row to the meetings index. Link notes, stills catalog, glossary. Don’t link the video.
Order of work (checklist)
- Probe video; extract wav; chunk; Deepgram via OpenRouter.
transcript.mdwith Speaker N:M + light Burra-class fixes.meeting-notes.mdfrom audio (decisions / actions).- Inclusive
screenshot-timestamps.json+.md. - Test one still’s rotation; extract all stills.
stills-catalog.md(screen vs missed).- Grep
bhgsite2026HTML for names the stills or ASR are unsure about. - Correct transcript; rewrite notes from catalog.
- Update glossary + meetings README (workshop disk).
- Stop. Discuss sensitive / internal-only spans with Emlyn. Write
published/notes and transcript. Then publish to BHG Site Working and commit if asked. - Leave video, temp audio, keys, and JPEGs out of the commit unless asked.
Worked examples
28 August 2026: 73 minutes, 185 grabs. Useful stills confirmed Preview URLs, Inc./plaques, Resources tiles, VIC price list, Goyder cemetery link text, Weddings duplicates, and the CHERRY/KIDD / GURNEY/KNAGGS / HINCH duplicate-block bug. Maps opening, Features, and Mongolata were almost all off-camera — those sections in the notes stay audio-only (plus HTML for Creek Street).
4 September 2026: four clips (~32 minutes of tape, with gaps) and five phone stills. The camera almost never found the monitor; the books, the 2017 booklet table, and the IREN card are why the stills pass still matters. Clip A talked through Resources publications and the workshop Useful Links inventory anyway.
21 September 2026: one 42-minute clip. Monitor for the first three minutes (BHG-005 ticket + Harvard Ask Barbara table); camera covered after that. Full tape in docs; working-site copy is published/.