elevenlabs.providers.sgit.ai / Example files, and the lexicon
Example files, and the lexicon
Files, not screenshots of files. Every one of these was written from the vendor's reference by a session that could not reach the API, and none has ever been executed. They are published because unrun code with a badge is more useful than no code — and because the first person to run one will know more than the people who wrote it.
Prose from the video vault at commit 7d1916aca5f3, 7 September 2026. When the vault moves ahead, this page is behind — and says so rather than guessing.
Badged unrun, and meant literally. written, not run verified 4–5 Sep 2026 Expect a wrong field name, a changed response shape, or an endpoint that has moved. If you run one and it works — or does not — the correction path is this site's repository, and the ledger row changes state with the date.
The scripts
| File | What it does | State |
|---|---|---|
00-smoke.sh | 00-smoke.sh — does the key work, what is the quota, and one sentence of audio. | written, not run |
align.sh | align.sh <audio-or-video> <script.txt> [out.json] — forced alignment of an existing render to its script. | written, not run |
stt.sh | stt.sh <audio-or-video> [out.json] — Scribe transcription with word timestamps (the QA gate, docs/05). | written, not run |
tts-timestamps.mjs | tts-timestamps.mjs "<text>" <voice_id> [basename] — one request to /with-timestamps, writes | written, not run |
tts.sh | tts.sh "<text>" <voice_id> [out.mp3] — plain generation, MP3 to a file. ELEVEN_MODEL and SPEED honoured. | written, not run |
voices.sh | voices.sh — list voices as: voice_id · name · category · accent · gender · use_case | written, not run |
Each expects a key in the environment and nothing else — no SDK, no npm install, no dependency you have to trust:
export ELEVENLABS_API_KEY=... # your own key, for one shell bash 00-smoke.sh # is the key alive, what is the quota, one sentence of audio bash voices.sh # voice_id · name · category · accent · gender · use case bash tts.sh "the text" <voice_id> out.mp3 node tts-timestamps.mjs "the text" <voice_id> /tmp/out # → .wav + .words.json + .srt bash align.sh render.mp4 script.txt # forced alignment of an existing video to its script bash stt.sh render.mp4 # Scribe transcription, for the QA diff
In the estate they came from, the key is not typed at all — it is opened from the vault's sealed config for the life of one command:
export ELEVENLABS_API_KEY=$(node tooling/scripts/vault-secrets.mjs open elevenlabs)
tts-timestamps.mjs is the interesting one: it asks for pcm_44100 so the bytes go straight into a Float32Array with no decoder, wraps them in a 44-byte RIFF header by hand, then does the characters → words → cues arithmetic and writes an SRT beside the WAV. It is 33 lines and has no dependencies. It has also never been run written, not run.
The lexicon
Download pronunciations.pls written, not run
A PLS lexicon for the estate's own vocabulary — sgit, sgit.ai, SG/Vault, SGraph, AIUC-1, SHA-256, llms.txt, OWASP, PT-BR, José, v0.1.29. Nobody publishes one of these, which is the only reason it is worth shipping: it is a starting point for anyone whose narration contains product names, initialisms and version numbers.
It is a hypothesis. It has never been uploaded, and not one alias in it has ever been heard aloud. Two facts govern how to use it vendor docs 5 Sep 2026:
- alias rules are text substitution before synthesis, and work on every model;
- phoneme rules (IPA or CMU Arpabet) are honoured only by
eleven_flash_v2andeleven_v3— other models ignore them silently, which is the failure mode to design against.
Start with aliases; move a word to a phoneme only when an alias still comes out wrong. The pronunciation lab exists to do exactly that comparison, and prints the result as markdown you can paste back into the file.
Why they matter more here than usual
In the pipeline these came from, the narration text is the caption text — the words the voice speaks are drawn on screen. So every pronunciation hack is visible: the scripts in that vault say sgit dot ai, A I U C one, S H A two five six, version zero point one point twenty-nine, and viewers read them in the caption band as typos.
A lexicon moves the hack off the screen: the text says sgit.ai, the voice says it right, and the caption band reads like prose for the first time. That is a change a viewer would notice before they noticed the voice.