Flux TTS pause markers: 8 per request, 500-3000 ms, and Sume scripts

A Flux TTS request can carry at most 8 inline pause markers, each from 500 to 3000 ms in 100 ms steps, and only on batch (REST) requests. That is a script-planning limit: a long narration with a pause after every paragraph will not fit. Sume's TTS request takes up to 20000 characters of transcript, and its sentence segmentation returns timings you can use to place silence yourself.

Sources: the Deepgram changelog entry of September 30, 2026 and Sume's API reference, read 2026-10-01.

What are the Flux pause limits?

How do I plan a script around 8 pauses?

Count your breaks before you write. Spend the eight markers on the pauses that carry meaning, such as before a reveal or after a price, and let punctuation handle the rest. If a script needs more, split it into several requests so each carries its own eight. Remember the speed cap: a pause in the text means speed cannot go above 1.15.

What does Sume give me for pacing?

Sume's schema has timestamps.words: true, which returns word-level timings on the completed job, and segmentation for sentence segments that are gapless: each segment ends where the next begins. The schema says segmentation requires timestamps.words: true. With those timings you can insert silence in your own edit rather than inside the request.

Speed is a [0.6, 1.5] multiplier in generation_config, and an optional emotion guide is available. The cited schema does not describe an inline pause marker, so this post does not claim one.

When is each approach the better fit?

Use inline markers when you want the silence generated with the voice and your script has few breaks. Use timings and an edit when you have many, or when the pauses depend on the picture. For a wider comparison of marker styles see adding pauses in text to speech.