Text to dialogue API: continuity between requests on Sume

Sume's text-to-speech request has no field for the text before or after a line, so continuity between separate requests is not something you pass in. Put the lines that must flow into one transcript, or render takes separately and join them with Timeline audio concat.

The ElevenLabs changelog entry for Sept 28 added previous_text and future_text (max 100 characters each) and previous_request_ids / next_request_ids (max 3) to Text-to-Dialogue. Sume's OpenAPI contract, read 2026-09-30, contains none of those names.

What can I do in one Sume request?

POST /v1/tts-1.0/generate accepts a transcript of up to 20000 characters. Optional timestamps.words with segmentation.mode: "sentence" returns word timings and gapless per-sentence segments[] (segment[i].end == segment[i+1].start); audio slices additionally need emit_audio and a wav or raw container. Only "sentence" is supported in v1. If the whole exchange fits in one transcript, one request keeps its delivery in one render.

How do I join separate takes without a gap?

Import the audio, then call POST /v1/timeline-1.0/audio with operation: "concat" and up to 20 ordered parts[]. The docs describe the join as sample-domain: no re-TTS and no silence at the seams. The call needs an Idempotency-Key, and the job is polled through /v1/jobs/:id/status and /result.

What can go wrong when joining?

Parts must share one channel layout, or the job fails with audio_parts_channel_mismatch. Keep wav output (the default) when the file will be joined again: mp3 re-adds priming padding at every edge. Joining preserves each take as rendered, so a delivery mismatch between two takes is not smoothed; it is only made gapless. Details are in Timeline audio.

Which one should I pick?

Use one transcript when the lines fit in 20000 characters and one voice setup. Use concat when takes are produced at different times or need individual retries. For per-line voices, see text to speech with multiple voices.