capability
verified July 2026
Spex-TTS — the engine.
One model serves every request on this site — the console, the specimens, the benchmarks, and your calls. This sheet is the engine itself: where it came from, what it publishes, and how a request steers it.
The signature — one grain, cast live
in-house sand engine · one shot
Spex, from speck — a single grain of sand. The material this site is built from, at its smallest unit: one grain, one engine, every voice on the wire.
The meter bank — three published figures, each worked in full on the sheet that owns it
first audio, ms
107 / 108
p50 / p95 on a single stream — server time on the production API, drawn against the one-second beat a caller notices.
methodology — the latency benchmark
languages, one cloned reference
23
One reference carries one identity across all 23; the inner tick is the eight that carry published, dated specimens.
specimens — the languages page
concurrent callers, one GPU
16
Zero failed requests at the measured edge; past it the API refuses fast with a retryable 503 rather than hanging a call.
the load benchmark
The datasheet on the product tour collects the full set; these three are the engine’s signature. Every number is measured on the production API, basis stated where it is worked.
Origin
Where it comes from
Spex-TTS was not built to demo synthesis. It grew up inside a production live-translation product speaking to thousands of people at once, where every turn had to land inside a conversation’s rhythm and a cloned identity had to survive crossing into languages its owner never spoke. The API exposes what that product forced into existence — nothing on these sheets started as a roadmap item.
Controls
Expression, per request
Delivery is not a deployment setting here — it is request fields. Six numeric controls (expressiveness, temperature, cfg_weight, speed, volume, seed) set intensity, melody, pacing, rate, gain, and repeatability; inline transcript controls — a spell-out tag, an emotion field, a pronunciation dictionary, timed breaks — steer the read itself. All of them ride every surface with no added latency. Ranges, defaults, and the expressive recipe are in the API reference at /docs/.
{ "transcript": "Your code is <spell>TKT4829XB</spell>.‣ a spell-out tag reads the code character by character <break time=\"600ms\"/> Read it back to me.",‣ a timed break, exactly where the read needs air "voice": {"mode": "id", "id": "gandr-dane"}, "emotion": "calm",‣ a tuned preset under the hood "expressiveness": 0.6,‣ explicit fields always win "cfg_weight": 0.4,‣ spacious pacing for the read-back "seed": 42‣ the same take, every render}See also
Related sheets.
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.
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.
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.
capability
3 tiers
Serving that degrades before it fails
Three serving tiers behind one endpoint — primary GPU fleet, warm spill, independent fallback — and a fast retryable 503 past absolute capacity.
Your script, this API, thirty minutes — and the recording leaves with you.
Hear it on your own script