> 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/12/ > > 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. # Singing and gemination In the text-to-speech lane: Amharic gemination (consonant lengthening) is phonemic but unwritten in the script, yet it takes more musical time, so text-setting for song cannot be derived from orthography alone. Singing synthesis for Amharic is rare, and verifying sung output by speech recognition is unreliable because singing distorts the formants ASR depends on. Stress is weak and not lexically contrastive, which helps.

× What it broke

Any plan to check synthesized Amharic song by running it back through ASR. The check would fail on good output as readily as bad.

✓ What it established

A property of the script worth knowing before any speech work: the written form under-specifies the spoken form in a way that matters for timing, and a native reader supplies what the orthography leaves out.
Evidence: build report §4.12 · 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/12/ ```

Every page here has a markdown twin; this page's is https://ethiopia-build.stoagen.com/problems/12/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 - This finding comes from a production lane (text-to-speech and song) adjacent to the corpus work and is included because it is a concrete Amharic-specific property the team recorded. No synthesis was produced. - Gemination, for a reader without the background: a consonant held for longer, as a distinct sound from the short one, so that two words can differ only by how long a consonant is held. In Amharic this difference changes meaning (it is phonemic) but the Fidel script does not mark it; a reader supplies it from knowing the word. This is the general shape of several problems on this site: the written form under-specifies the spoken form, and the missing information lives in a native speaker. - Why it matters beyond song: any text-to-speech system working from Fidel has to guess gemination, and a transcription system working from audio hears it but has nowhere standard to write it. So a round trip (speech to text to speech) loses it unless the pipeline carries it as annotation outside the script. The record does not propose a notation; it records the gap. - Why ASR cannot verify sung output: singing holds vowels and moves pitch, which distorts the formants that speech recognition uses to identify vowels; a transcript of sung Amharic from a speech model is unreliable even when the singing is correct. The check fails in both directions. The only verification the record considers valid is a native listener, which is the same conclusion as problem 02 for text. - The helpful property: Amharic word stress is weak and does not distinguish words, so text-setting does not have to preserve a stress pattern, only gemination and vowel length. - This page is the least connected to the corpus pipeline and is kept because the list is meant to be complete: every concrete Amharic problem the team met, wherever it met it. - Related: 10 (ASR), 02 (human review as the only verification), 13 (script limits). - Source: build report §4.12.