capability
verified July 2026
Voice cloning from ten seconds, in the request.
No enrollment pipeline, no per-voice fee, no waiting for a training job. The reference clip is a request parameter, and the identity it carries survives translation.
The tape — voice specimens, pre-rendered on the production API
generated 2026-07-16
Leader · The whole enrollment
Five to ten seconds of one consenting speaker, up to ~1.5 MB of WAV. The clip becomes a request field; references are fingerprinted and cached, so the first call pays the cloning work and every reuse skips it.
import { readFileSync } from "node:fs"
// ten seconds of one consenting speaker — the whole enrollment
const wav_b64 = readFileSync("reference.wav").toString("base64")
const body = {
transcript: "Same identity, any of 23 languages.",
language: "de", // a language the
voice: { mode: "clone", wav_b64 }, // reference never spoke
}
// fingerprinted + cached: the first request pays the cloning
// work, every reuse of the same clip skips it
Scale
each strip is printed at its clip’s real length on this scale
Side A · the money desks
Scripts that cannot afford a stumble: a fraud hold, a payment plan, a first notice of loss.
Leo · English
financial services · fraud hold
“I've placed a temporary hold on the card ending four two one nine after two charges that don't fit your pattern. If those were you, I can lift it right now. If not, I'll have a new card on its way before we hang up.”
Dane · English
collections · payment plan
“Your balance is about thirty days past due, and I'd much rather work it out with you than let it escalate. I can split it into two payments — would the first one landing this Friday work for you?”
Lewis · English
insurance · first notice of loss
“I've opened the claim for the water damage and assigned it to an adjuster — they'll reach out within one business day. Hold on to the photos you took; they'll move everything along faster.”
Side B · the front desks
The appointment moved, the refund landed, the table is booked — the everyday floor of a support line.
Jenny · English
healthcare · appointment line
“You're all set for Thursday at ten with Doctor Okafor. I've texted the address and the two intake forms — reply here anytime and I'll move the appointment if something comes up.”
Mia · English
retail · returns desk
“Good news — your refund went through this morning, and the confirmation number is already in your inbox. It usually settles back to your card within two business days. Anything else I can take care of?”
Ava · English
hospitality · front desk
“Good morning, and thank you for calling the front desk. I can check you in, arrange a late checkout, or book you a table for tonight — what would help you most right now?”
Side C · operations
Dispatch, field scheduling, and the hard words on the first try.
Dane · English
logistics · dispatch
“Driver twelve is eight minutes out with two pallets, and dock four is clear. I've dropped the manifest in your inbox so you can check it before they pull in.”
Leo · English
field service · scheduling
“Your technician is booked for Tuesday between nine and eleven. I'll send a text when they're about thirty minutes out, so you're not stuck waiting around all morning.”
Dane · English
hard words, first try
“Ondansetron twice daily; the subpoena arrived Thursday; the charcuterie is from Worcestershire.”
Side D · one voice, crossing languages
The hard case is a clone speaking a language its reference never spoke — the identity has to survive with none of the original phonemes to lean on. That case held up in blind listening panels; one reference carries one identity across 23 languages, and six multilingual stock voices ship ready-made when you do not need a clone.
Jenny · English
one voice, every language
“Your driver is eight minutes away and dock four is clear — the paperwork is already in your inbox.”
Jenny · German
one voice, every language
“Ihr Fahrer ist acht Minuten entfernt und Dock vier ist frei — die Papiere liegen bereits in Ihrem Postfach.”
Jenny · Arabic
one voice, every language
“سائقك على بعد ثماني دقائق والرصيف رقم أربعة جاهز — والمستندات موجودة بالفعل في بريدك.”
Jenny · Japanese
one voice, every language
“ドライバーはあと八分で到着し、四番ドックは空いています。書類はすでにメールでお送りしました。”
Run-out · Provenance, stated
- Every synthesized sample carries an inaudible watermark, applied at generation on every serving tier.
- Customer reference audio never enters a training set.
- Usage is per-key and auditable at GET /v1/usage — the full policy sheet is at provenance and privacy.
01
Notes — an engineer's checklist
01What makes a good reference clip?
Five to ten seconds of one speaker talking naturally in a quiet room. Clean beats long — the cap is ~1.5 MB of WAV, about 30 seconds, and extra length past ten seconds buys little.
02Do I pay per cloned voice?
No. There is no per-voice fee and no voice limit — a voice is a request parameter. A platform can carry one clone per end customer without a pricing conversation.
03Does the same reference sound the same tomorrow?
The identity holds — references are fingerprinted, so the same clip maps to the same cached voice. The engine itself is expressive rather than deterministic: two renders of the same sentence differ the way two takes differ, unless you pin a seed, which makes the same text, voice, and parameters reproduce the same render.
04Can a clone speak languages the speaker does not?
Yes — that is the cross-language property the engine was built around. One English reference speaks all 23 supported languages, and the blind-panel result is that the identity survives the crossing.
See also
Related sheets.
glossary
10 s
Instant voice cloning
Instant cloning builds a voice from a ten-second reference at request time — no training job. What zero-shot means, and how identity survives translation.
glossary
~1.5 MB
Voice cloning reference audio
The reference is the clip a zero-shot clone is built from: about ten seconds of one clean, consenting speaker, a ~1.5 MB cap, fingerprinted and cached.
use case
23
TTS for live voice translation
Spex-TTS grew out of a production live-translation product. One cloned identity across 23 languages, 107 ms first audio, measured on the production API.
capability
never
Provenance and privacy, applied at generation
The policy surface of the engine: an inaudible watermark on every sample, reference audio that never trains, and per-key audit trails at GET /v1/usage.
Your script, this API, thirty minutes — and the recording leaves with you.
Hear it on your own script