Start a speech-to-text transcription. Asynchronous — poll for the result.

Multipart form: {file, model, language_hints?, enable_speaker_diarization?} — the last two carry Soniox's own field names. Repeat language_hints once per code (default en, ko). Returns: {operation_id, status: "processing", model, duration_seconds, credits_charged}. Poll GET /v1/gateway/audio/transcriptions/{operation_id}/.

Billing is settled here, off the duration probed from the uploaded file — polling never charges again. A file over the 2-hour cap, or one whose container carries no readable duration, is a 400 before any provider work.