elevenlabs.providers.sgit.ai / experiments / Voice-settings sweep
Voice-settings sweep
Four sliders, and the vendor's guidance on them is qualitative. This lab sweeps one at a time, holds everything else constant, fixes the seed, and makes you write down what you heard — which is the only part of this that is not arithmetic.
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 line, and the knob to sweep
2 · The sweep
Held constant: everything not being swept. That is the whole method — one knob at a time, same seed where you set one, same text, same voice. Listen twice before you write anything down.
| Value | Latency | Spoken length | Bytes | Listen | Your note |
|---|
3 · Log
What this tests
speedis the setting that sent us here in the first place. Its documented range is 0.7–1.2, default 1.0, with quality costs at the extremes vendor docs 5 Sep 2026. The incumbent provider has no equivalent, and the absence cost a portrait cut a third of its script instead of 15% of its pace measured 2 Sep 2026. The measurable question: how much time does 1.15× actually save on your script, and at what point does it stop sounding like a person? The "spoken length across the sweep" line answers the first half; your ears answer the second.stabilitytrades expressiveness against consistency. Oneleven_v3the vendor frames it as three modes — creative, natural, robust — rather than a continuum written, not run. For narration of somebody else's compliance standard, hallucination is not a risk worth taking, so the interesting range is the top half.stylecosts latency and is worth 0 for most narration. This lab is where you confirm that rather than assume it.similarity_boostmatters most for cloned voices; on a library voice the effect is usually small, which is itself worth measuring once.
Method
One knob, five values, everything else pinned, seed fixed at 42 so the read is as repeatable as the vendor allows — the seed is documented as best-effort, which the text-handling lab tests directly.
Latency is reported but is not the point here: a sweep is sequential, so the numbers include whatever your connection was doing at the time. The two columns that matter are spoken length — a real, comparable measurement — and your note, which is the only column that captures quality.
What it costs
Five values of a 90-character line is about $0.045 on v3, half that on flash vendor docs 5 Sep 2026. A full sweep of all four settings is under a quarter of a dollar, which is less than the cost of arguing about it.