Sume TTS segmentation 400: segmentation_requires_word_timestamps

A TTS request with segmentation but without timestamps.words: true is refused with a message naming segmentation_requires_word_timestamps. Add "timestamps": { "words": true } beside the segmentation object and send it again.

A valid pair

The two blocks work together: word timestamps are the input to sentence segmentation. segmentation.mode accepts only sentence.

{
  "model": "sonic-3.6",
  "transcript": "First sentence. Second sentence.",
  "voice": { "id": "VOICE_UUID" },
  "timestamps": { "words": true },
  "segmentation": { "mode": "sentence", "emit_audio": false }
}

Segmentation options

boundary_lead_ms is an integer from 0 to 500 and defaults to 70. emit_audio defaults to true and needs a wav or raw output container, so set it to false when you keep the default mp3 and only need the times.

Why a 400 will not fix itself

A validation refusal does not change on retry. Correct the body before resubmitting, with a new idempotency key if your client treats the failed attempt as final. The boundary lead post explains what the 70 ms lead does to cut points.