elevenlabs.providers.sgit.ai / experiments / Pronunciation lab — lexicons, and the test nobody has run
Pronunciation lab — lexicons, and the test nobody has run
In this pipeline the narration text is the caption text, so every pronunciation hack is on screen where viewers read it as a typo. A dictionary moves the hack off the screen. Whether these particular rules work has never been tested by anybody.
Your own full key, in your own browser, bounded only by your plan's monthly quota. That is pattern 0 with a ceiling: the narrow case where it is defensible — the key's owner, testing their own key, on their own device — and not a pattern to publish. No key ships in this page: the field below is empty until you fill it.
ElevenLabs offers no per-key spend limit for text to speech, so nothing here caps what a leaked key could cost you except the account's own quota. Use a key scoped in the dashboard to the endpoints this lab needs, and forget it when you are done. The version of this page with no key box at all is pattern three, and it does not exist yet.
1 · The lexicon
| Grapheme | Type | Alias or phoneme |
|---|
alias rules are text substitution and work on every model. phoneme rules (IPA)
are honoured only by eleven_flash_v2 and eleven_v3 — every other model ignores them
silently, which is the failure this lab exists to make visible.
2 · Upload it, and pin a version
3 · A/B — the same text, with and without the dictionary
Without the dictionary
With it
4 · Write down what you heard
This is the open item the whole site keeps pointing at. Mark each name, then copy the report into the repository or the vault — it is the difference between a hypothesis and a lexicon.
| Token | Without | With | Should sound like |
|---|
5 · Log
The problem, concretely
Four script files in the source vault contain, verbatim: sgit dot ai, A I U C one, S H A two five six, llms dot txt, version zero point one point twenty-nine. Each of those is a workaround for a text-to-speech engine, and each of them is drawn on screen under the picture, where a viewer reads it as a mistake.
With a dictionary the script says sgit.ai and the voice still says it correctly. That is a change a viewer notices before they notice the voice.
The two rule types, and the trap
| Rule | What it does | Works on |
|---|---|---|
| alias | text substitution before synthesis: sgit.ai → ess-git dot A I | every model |
| phoneme | exact pronunciation in IPA or CMU Arpabet | eleven_flash_v2 and eleven_v3 only |
Other models ignore phoneme rules silently vendor docs 5 Sep 2026 — no error, no warning, just the old pronunciation. That is the trap this lab is built to expose: generate with a phoneme rule on eleven_multilingual_v2 and listen to nothing happen. Up to three dictionaries may be attached to a request, and each request pins a version, so a dictionary edit does not silently change an old render.
The open item
The lexicon shipped on this site — pronunciations.pls — is a hypothesis written, not run. It has never been uploaded and not one alias in it has been heard. Neither has the prior question: which names the model gets wrong without any dictionary at all written, not run.
Section 4 is the point of the page. Generate both sides, listen, mark each token right or wrong, copy the markdown, and the site gains its first piece of evidence about pronunciation. The default text is the names sample; it costs about $0.02 for both sides on v3 vendor docs 5 Sep 2026.
A shortcut worth knowing
On eleven_v3 you can write IPA inline in the text between slashes — /ˈsɪdʒɪt/ — which is a fast way to find a transcription that works before committing it to a dictionary version written, not run. Find it inline, then move it into a rule; do not ship inline phonetics in narration text, because that text is what appears on screen.