is_final_audio_for_turn vs Sume TTS sentence boundaries: wav or mp3

is_final_audio_for_turn is an ElevenLabs streaming marker; Sume TTS jobs have no such flag. Sume gives you sentence boundaries instead: with timestamps.words and segmentation on, you get gapless segment timings, and sliced per-segment audio only when the output is wav or raw.

The ElevenLabs changelog for 2026-09-21 (read 2026-10-01) says the marker is now emitted after every buffered byte for the turn, including MP3 and Opus, so clients can use it as an exact turn boundary without switching to PCM.

What does the ElevenLabs marker tell a client?

It marks the end of a turn's audio in a multi-context stream. The changelog's point is that the marker no longer depends on PCM output.

How does Sume mark where a sentence ends?

The TTS schema describes optional sentence segmentation that requires timestamps.words: true. It returns gapless segments[], where the end of one segment equals the start of the next. boundary_lead_ms is the number of milliseconds after the last word of a sentence before the cut, default 70; the next segment absorbs the pause.

Which format should I request?

The schema states: "For mp3, timings are returned without segment audio_url." If you need playable per-sentence files, ask for wav. If you only need timings to drive captions or a cut list, mp3 is enough. See mp3 or wav for TTS for the format trade-off.

Is this a streaming boundary?

No. A Sume TTS job completes and then returns its result; the boundaries are in the finished result, not in a live byte stream. If your player needs a live end-of-turn signal, that is the ElevenLabs behavior the changelog describes.