POST /v1/audio/speech

Generate audio from text (unit-billed, chars)

Specialty route (route_type: speech); per-alias input caps set 2026-09-09. The response is BINARY audio via chunked transfer: bytes leave the gateway as they arrive from the model (flush per read — no idle gap past load-balancer walls), never a Content-Length. Content-Type forwards from the model (default audio/mpeg); no other header from the model crosses. No stream param — chunked transfer is the only mode. Input is length-capped per alias where the model does not cap itself (400 invalid_param param=input, refused before any connection is made, unbilled; values in the gateway's models.yaml, sized so buffered synthesis fits the model host's ~55-60 s no-byte edge await and the gateway's per-alias first-byte wall: chatterbox-tts 400 (measured 2026-09-09: 0.085 s/char buffered synthesis — 600 chars reached 53.6 s to first byte; the stub-era 20000 is gone) · kokoro-tts 3500 (measured 0.0149 s/char — ~52.8 s at cap; 4000 would run ~60 s, an edge kill) · pocket-tts 2500 (measured 0.0188 s/char — ~48 s at cap) · audio8-tts 250 runes (measured 2026-09-09: ~0.22 s/char measured × 250 = ~56 s under the edge kill — the old 400 derived from a ~100 s wall that does not exist)). The model synthesizes BEFORE the first byte crosses — first-byte latency of tens of seconds is normal at the caps (chatterbox ~41 s at 400 chars); text longer than a cap is the client's to chunk (one request per chunk, concatenated at playback). Metering: unit_type='chars' = rune count of input, counted at the gateway. Idempotency-Key is accepted with NO replay semantics on this route (§8 route-scope) — a retrying client re-executes and re-bills. Client contract: do not set client read timeouts below the expected generation time.

curl -X POST https://api.ironstratum.com/v1/audio/speech \
  -H "Authorization: Bearer $IRONSTRATUM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"<model>","input":"<input>","voice":"<voice>"}'