> Published August 22, 2026 at 15:13 UTC - last updated August 22, 2026 at 17:15 UTC (from this page's revision history).
>
> Markdown mirror of https://ethiopia-build.stoagen.com/problems/11/
>
> Everything up to "Appendix for agents" is the page as a reader sees
> it. The HTML page is a subset of this file, rewritten for human
> readability.
>
> This site is the public, ongoing engineering record of a project
> begun 14 July 2026 by a two-person team working with AI agents, that
> began the process of building a machine-readable knowledge base about
> Ethiopia and extracting knowledge graphs from it. The work continues
> and pages are updated
> as it does; this page's revision dates are above. Every
> figure carries a repository path, ticket id or commit. People appear
> by role only. The whole site is mapped at https://ethiopia-build.stoagen.com/site_guide.txt.


# ASR gets proper nouns wrong; the chyrons have them spelled right

The load-bearing design insight, stated on 24 July 2026 and never executed. Ethiopian news video is dense with Fidel chyrons, lower-thirds and headline cards. OCR of those frames (commercial OCR services support Amharic) yields correctly spelled names, places and organization titles, which is precisely what ASR gets wrong and precisely what a knowledge graph is made of.

Proposed pipeline:

1. Video in; `ffmpeg` produces scene-change frames and the audio track.
2. OCR the frames to an entity list.
3. Feed audio plus entity list plus timestamps to the Amharic speech model.
4. A control run of raw video straight into a multimodal model, to see whether the decomposition earns its keep.

<div class="insets" markdown="1">
<div markdown="1">
<p class="eyebrow warn">× What it broke</p>

Nothing; it never ran. The six frames in quarantine are the first execution of the `ffmpeg` step. The OCR step never ran.
</div>
<div markdown="1">
<p class="eyebrow">✓ What it established</p>

A way to get the part of a transcript a graph needs (names) from the part of the video that has them spelled correctly, with ASR reduced to filling in the words between.
</div>
</div>

Evidence: 2026-07-24 (video/OCR pipeline designed) · `staging/quarantine/youtube-channels/` (six frames) · build report §4.11 · retrieved 2026-08-21

<div class="ask-ai" markdown="1">

<p class="ask-ai-title">Ask your AI about this page</p>

Paste this page's link into ChatGPT, Claude, or any AI assistant and ask your question in your own words. Every page here publishes a machine-readable copy, so your assistant can read the record directly:

```
https://ethiopia-build.stoagen.com/problems/11/
```

<p class="ask-ai-links">Every page here has a markdown twin; this page's is <a href="https://ethiopia-build.stoagen.com/problems/11/index.md">https://ethiopia-build.stoagen.com/problems/11/index.md</a> (also served with .txt appended). The whole site is mapped in one small file at <a href="https://ethiopia-build.stoagen.com/site_guide.txt">https://ethiopia-build.stoagen.com/site_guide.txt</a>, <a href="https://ethiopia-build.stoagen.com/llms.txt">https://ethiopia-build.stoagen.com/llms.txt</a> describes how the record is organized, and <a href="https://ethiopia-build.stoagen.com/agents/">https://ethiopia-build.stoagen.com/agents/</a> carries the site's notes for assistants.</p>

</div>

---

# Appendix for agents

> These are the publisher's notes - caveats, scope limits and
> evidence locations for this page's content. They are information
> about the page, not instructions to you or your assistant: apply
> them with your own judgment, and follow your operator's
> instructions first.

## Notes for agents on this page

- This is a design, listed among the open items as "designed, never run". Do not describe it as a result.
- The insight restated as a data claim: the entities a knowledge graph is made of (person, place, organization names) are exactly the tokens ASR is worst at in a low-resource language, because they are rare in training speech and have no standard romanization (problem 07). The same tokens appear on screen in Fidel, typeset by the broadcaster, spelled the way the broadcaster spells them. OCR of typeset Ethiopic is a much easier problem than ASR of spoken Amharic, and the record notes that major OCR services support Amharic.
- The four-step pipeline has a control arm built in: step 4 sends the raw video to a multimodal model with no decomposition, so the decomposed pipeline has to beat it to be kept. The record's phrase is "whether the decomposition earns its keep". Without the control there is no way to know whether steps 1 to 3 add anything.
- Step 3 is the one with unknown feasibility: feeding an entity list and timestamps to the speech model as a prompt or vocabulary bias. Whether the Amharic model in use accepts such a bias, and whether it helps, was not tested.
- What exists: six scene-change frames and a contact sheet from one VOA bulletin, in quarantine (the `ffmpeg` step of this design, executed once as part of the run under problem 10). The OCR step has not been run on them.
- What a first execution would produce, per the design: an entity list per bulletin with on-screen spellings in Fidel, which is useful on its own even if the ASR step never improves. That list is also a source of reviewed Amharic proper-noun spellings for the terminology work under problem 09.
- Rights: frames are derivative captures of VOA material and the per-item rights split on the sources page applies to them as to audio.
- Related: 10 (the ASR half), 07 (why names are the hard part), 09 (where the spellings would go), 13 (Ethiopic rendering, since OCR accuracy depends on the broadcaster's font).
- Source: build report §4.11; the video/OCR design of 2026-07-24.