Poem reading video: slow TTS, an emotion string and a ducked music bed

Set generation_config.speed to 0.6, the slowest the TTS router allows, add an emotion string, and generate the music bed separately. Then render with a soundtrack that has duck_db set, so the bed drops under the voice. The three jobs cost a few cents of speech plus $0.125 for music.

Step 1: the slow reading

POST /v1/tts-router/generate requires model. The generation_config object takes speed from 0.6 to 1.5, volume from 0.5 to 2 and an emotion string of up to 64 characters. The voice comes from voice.id, which must be a UUID or a voi_ id, or from avatar_id or avatar_handle.

curl -X POST https://api.sume.com/v1/tts-router/generate \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: poem-voice-001" \
  -d '{
    "model": "sonic-3.6",
    "transcript": "I wandered lonely as a cloud, that floats on high o'\''er vales and hills.",
    "voice": { "id": "'"$SUME_VOICE_ID"'" },
    "generation_config": { "speed": 0.6, "emotion": "calm" }
  }'

Step 2: the bed and the mix

Ask the Music Router for a quiet, sparse instrumental, for example Solo piano, slow, sparse, instrumental, no vocals. A 1-minute track. Length is steered in the prompt. In the Timeline 1.0 render, pass the speech as audio.url and the music as soundtrack.

Checks before you ship

Listen at 0.6 speed, because very slow delivery can sound stilted on short lines, and raise it a notch if it does. Keep duck_db modest. A poem needs the bed audible in the gaps. If the song is shorter than the spoken track, set loop on the soundtrack and a fade_out_seconds of a few seconds. The bedtime story post uses the same structure.