> 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/10/ > > 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. # Speech: the one ASR attempt died on a missing DLL On 23 July 2026 a two-pass `faster-whisper` job (transcribe in Amharic, then translate to English, beam size 5, voice activity detection on) was run on six audio slices from a VOA bulletin, with six extracted video frames and a contact sheet beside them. The progress log stops at "transcribing, language am". The error log ends: ``` RuntimeError: Library cublas64_12.dll is not found or cannot be loaded ``` The model had already been downgraded from `large-v3` to `medium`, and the model cache ran in degraded mode because the machine did not support symlinks.

× What it broke

Zero Amharic ASR output exists anywhere in the programme. The attempt was not repeated.

✓ What it established

The script was already honest about what it would have produced: its markdown templates carry a fixed notice that the output is an automated program transcript, not a VOA-supplied one. That labelling rule survives in the evidence model.
Evidence: `ethiopia-program/staging/quarantine/youtube-channels/` · 2026-07-23 · build report §4.10 · retrieved 2026-08-21

Ask your AI about this page

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/10/ ```

Every page here has a markdown twin; this page's is https://ethiopia-build.stoagen.com/problems/10/index.md (also served with .txt appended). The whole site is mapped in one small file at https://ethiopia-build.stoagen.com/site_guide.txt, https://ethiopia-build.stoagen.com/llms.txt describes how the record is organized, and https://ethiopia-build.stoagen.com/agents/ carries the site's notes for assistants.

--- # 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 - The failure is an environment failure (a missing CUDA runtime library on the machine), not a finding about Amharic speech recognition quality. No quality measurement exists because no output exists. - What was set up, so the next attempt does not repeat it: `faster-whisper` two-pass (language `am`, task transcribe, then task translate to English), beam size 5, VAD filter on, model `medium` (downgraded from `large-v3`), six audio slices cut from one VOA Amharic bulletin, six scene frames and a contact sheet extracted alongside by `ffmpeg` (the first and only execution of the step that problem 11 depends on). The error names `cublas64_12.dll`, the CUDA 12 BLAS runtime; the model cache ran degraded because the machine did not support symlinks, which is a Windows-privilege issue separate from the CUDA one. - Two things that were already right: the job was honest about its output class (the markdown templates carried a fixed notice that the result would be an automated program transcript, not a VOA-supplied one, which is the evidence model's layer four), and the input was sliced, so a successful run would have produced per-slice output that could be checked against the chyrons. - Expectations for a successful run, from the record: Whisper-family models list Amharic as supported, but the Common Voice audit (sources/speech) shows under two validated hours of public Amharic speech, so support and quality are different claims, exactly as for translation under problem 02. The record's position is that proper nouns are the expected weak point (problem 11), and that nothing should be reported as Amharic ASR quality until measured against a hand-checked transcript, of which none exists. - A successful run would also need the rights determination from the sources page: VOA-produced material is US government work; agency material VOA carries is not. The sampled bulletin's per-item status was not determined. - Open item: "Amharic ASR: zero output exists". It stays open until a run completes and is measured. - Related: 11 (the chyron OCR design this run was the front half of), 12 (why ASR is also unreliable for verifying sung output), 01 (the measurement gap). - Source: build report §4.10; `ethiopia-program/staging/quarantine/youtube-channels/`; 2026-07-23.