Text in. Voice out,
with the clock attached.

No, and we would rather say so here than have you find out. No speech engine returns word timings, so Speaker derives them one of two ways and every response names which in `timings_source`. Where an engine returns CHARACTER-level alignment, words are an exact reduction over it: group on whitespace, take the first character's start and the last character's end. Where an engine returns audio alone, Speaker transcribes its own output and aligns the transcript's word boundaries back onto your text — the clock comes from the transcript, the words are always yours.

You are not charged. A synthesis that returns audio but no usable word timings settles at zero credits and refunds in full, and on the keyless rail it is refused before the payment is settled. This product is voice with the clock attached; audio on its own is not what you asked for, and it was not your fault, so it is the house's loss.

Because `model` names a quality rung, not an engine's internal model. Speaker runs a chain of speech engines and falls through to the next when one refuses work, so a vocabulary made of one supplier's model names would be a promise the product could not keep the moment that supplier was down. Each engine maps Fast, Balanced and Rich onto its own models, and every engine's Rich is genuinely more expensive on that engine's own rate card. The response reports what actually ran, in `engine_model`.

Because that is how the cost behaves. Speech bills per character upstream, so a 4,000-character script genuinely costs forty times a 100-character one — and a flat per-call price is wrong at both ends: too expensive for a one-line reply, sold below cost for anything long. Fast and Balanced are 10 credits per 1,000 characters; Rich is 20, because every engine's own rate card charges roughly twice as much for the model it maps Rich onto. The price follows the rung you ask for and never the engine that serves it.

Send a pronunciations map: {"Ounie": "Oonie"}. It is applied to your text before synthesis, so the engine reads the respelling — and the word timings you get back carry YOUR spelling, not ours, so the captions are publishable as they arrive. That holds on every engine: where the timings come from a transcript, they are aligned onto your text in the first place, so the transcript mishearing the name (ours came back with "Oni") cannot reach you. A respelling has to be a single word; two tokens where you wrote one would put the timings out of step with your script, so that request is refused instead.

Your run is: the text, the voice, the timings and the audio, so you can list your history, re-read a run and download the audio again without paying twice. It is scoped to your account and nothing else. The audio object itself is private — every download is a signature minted for that request and expiring on its own, never a public URL.

5,000 characters, roughly five minutes of speech. The engine allows more, but this is a synchronous endpoint: the whole synthesis has to finish inside one request, and a limit the request timeout would break anyway is not a real capability. Longer scripts are several calls, and the word timings make them straightforward to stitch back onto one timeline.

Yes. The x402 endpoint takes USDC on Base with no account and no key. POST your text unpaid and the 402 carries the price for that exact text; sign, retry, and it verifies the signature, synthesises, and settles only once the audio and its timings both exist. An on-chain settlement is final, so a non-answer is refused before the money moves rather than refunded after.

You get an immediate 503 and nothing is charged — no credits held, no payment quoted, no signature asked for. A scheduled probe checks the engine on its own, so nobody's request pays for the discovery, and the moment it recovers the app starts working again with no deploy.