TTS speed: slow, normal, fast or generation_config.speed 0.6 to 1.5?

Use generation_config.speed, a number from 0.6 to 1.5, to change how fast Sume TTS speaks. The older top-level speed field, with values slow, normal or fast, is marked deprecated in the OpenAPI document, which says to prefer generation_config.speed. The same object also takes volume (0.5 to 2.0) and emotion, a string of 1 to 64 characters.

What are the exact ranges?

From the request schema for TTS 1.0 and the TTS Router: generation_config.volume is a multiplier in [0.5, 2.0]; generation_config.speed is a multiplier in [0.6, 1.5]; generation_config.emotion is a string, minimum length 1 and maximum 64. The object does not allow extra keys, so a misspelled field is refused instead of ignored.

Cartesia's Sonic 3.6 changelog entry (read 2026-10-02) says speed and volume controls behave on 3.6 as they do on Sonic 3.5, so a move to the newer model does not need new values.

Which should I send?

Send the number. It gives you steps between slow and fast, and it is the one the docs point at. Take 1.0 as your baseline, then adjust in small steps, such as 0.9 for a calmer read or 1.1 for a tighter one, and listen. Do not send both fields on one request unless you have a reason.

Your result will tell you what was used. A finished job records generation_config and speed, each null when you did not send it, as described in Jobs and results.

curl -X POST https://api.sume.com/v1/tts-1.0/generate \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: speed-090-001" \
  -d '{
    "transcript": "Take a slow breath in, and out.",
    "avatar_handle": "@your_avatar",
    "language": "en",
    "generation_config": { "speed": 0.9, "volume": 1.0, "emotion": "calm" }
  }'

Does a slower speed change the cost?

TTS usage counts the characters of the transcript, spaces and punctuation included, up to 20000. Speed does not change the character count. It does change the length of the audio, and synthesized audio over 1200 seconds fails with tts_duration_exceeded, so a long script at 0.6 can reach that limit before the character cap.