use case
verified July 2026
TTS for live voice translation.
This is the workload Spex-TTS was built inside. Live translation stacks two hard problems: the latency budget of a conversation, and a voice identity that must survive crossing into a language its owner never spoke.
01
Why translation is the hard case
A translated turn runs the whole pipeline twice-shifted: transcribe the source, translate, then synthesize in the target language — and the listener still expects conversational rhythm. Every stage pays its latency twice over the round trip, which is why the synthesis leg has to sit near 100 ms, not near 500.
Identity is the second problem. If the German side of the call does not sound like the English speaker, the translation feels like a third person in the room. In blind listening panels, our clone speaking a language the reference never spoke kept its owner’s identity — the property live translation depends on.
02
One identity, four of its languages
The same sentence, the same cloned reference, four scripts — dated production output. The other nineteen languages run live on a demo call.
Audio — One clone crossing languages — same reference, dated specimens
generated 2026-07-16
English
“Your driver is eight minutes away and dock four is clear — the paperwork is already in your inbox.”
Jenny · 5.5 s · one voice, every language
German
“Ihr Fahrer ist acht Minuten entfernt und Dock vier ist frei — die Papiere liegen bereits in Ihrem Postfach.”
Jenny · 6.4 s · one voice, every language
Arabic
“سائقك على بعد ثماني دقائق والرصيف رقم أربعة جاهز — والمستندات موجودة بالفعل في بريدك.”
Jenny · 7.6 s · one voice, every language
Japanese
“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。”
Jenny · 4.9 s · one voice, every language
03
What translation demands of synthesis
- One cloned identity across 23 languages, from one ten-second reference — no per-language training.
- A persistent WebSocket per call, an utterance per turn, binary PCM streaming back — the round trip pays the budget twice, so the socket setup is paid once.
- Unmetered characters: translation doubles the synthesized volume, and a meter doubles the bill for it.
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.
capability
0
Voice cloning from ten seconds, in the request
Zero-shot voice cloning: a ten-second reference rides inside each request, no training job, and the identity holds across 23 languages.
use case
8,000 min
TTS for AI customer-support agents
The latency budget and cost arithmetic for a voice support desk: why first audio decides whether callers interrupt, and what 8,000 minutes a month costs.
benchmark
5 rules
How we measure TTS latency
The methodology behind every figure on this site: server-time TTFA on the production API under call-shaped load, percentile pairs, published conditions.
Thirty minutes, your production script, the live latency readout — measured in front of you.
Bring this use case to the call