Flux controls are English-only at launch. Pronunciation controls apply to Flux batch REST, Flux WebSocket, and Aura-2 /v1/speak. Pause controls apply to Flux batch REST only.
Use client.speak.v2.audio.generate() for complete text. A pause is written as \{pause:500ms\} through \{pause:3000ms\} in 100 ms increments. A request accepts at most eight pauses. TextBuilder emits and validates this Flux batch syntax; its from_ssml() converter raises ValueError for out-of-range breaks, and its helpers reject mixed pronunciation and pause controls before sending a request.
Pronunciation uses \{"word": "...", "pronounce": "<IPA>"\} and is Early Access. A pronunciation cannot be combined with any pause or with speed other than 1.0. When a pause is present, speed cannot exceed 1.15. Invalid combinations return a batch error, including CONTROL_COMBINATION_INVALID or PAUSE_SPEED_CAP_EXCEEDED.
Read batch control outcomes from raw response headers: dg-pronunciations-applied, dg-breaks-applied, and dg-warnings. Warnings report best-effort pronunciation handling.
Use client.speak.v2.connect() for streaming Flux synthesis. WebSocket turns support pronunciation controls only. A pause marker such as \{pause:500ms\}, or a pronunciation marker such as \{"word": "...", "pronounce": "<IPA>"\} combined with non-default speed, closes the connection with DATA-0002.
Configure changes are buffered. Setting speed while a buffered turn contains pronunciation returns ConfigureFailure with CONTROL_COMBINATION_INVALID; flush that turn before changing speed.
SpeechMetadata.controls_applied reports pronunciation, pause, and warning counters. Warning messages report the associated warning code and description. WebSocket pauses are unsupported, so breaks_applied is always zero.
TextBuilder validates Flux batch constraints. Its pronunciation output is usable where pronunciation controls are supported, but do not use its pause output with the Flux WebSocket or assume Flux pause limits apply to non-Flux TTS endpoints.