Cartesia Emotion and SSML Controls in LyricWinter

Cartesia Sonic 3.5 can infer emotional subtext from the transcript, but a generated LyricWinter clip usually contains only one line. LyricWinter therefore looks at nearby detected dialogue, then passes Cartesia native controls when the surrounding scene gives useful delivery context.^[1]

The important difference from OpenAI TTS is the control shape. OpenAI accepts natural-language delivery instructions. Cartesia accepts documented controls: an emotion value, a speed multiplier, and a volume multiplier. LyricWinter maps the same nearby story context into those Cartesia-specific fields.

Automatic Cartesia Guidance

When a voice uses cartesia-sonic-3.5, LyricWinter can analyze up to 10 lines before and 10 lines after the current clip. If the context clearly suggests a performance direction, it sends Cartesia a generation_config with emotion, speed, and/or volume.

{
  "generation_config": {
    "emotion": "anxious",
    "speed": 1.08,
    "volume": 0.9
  }
}

If the scene is ambiguous or the line already works from the text alone, LyricWinter does not need to send a control. Cartesia's docs also warn that emotion guidance works best when it matches the transcript, so the system avoids forcing mismatched emotions.^[1]

Manual SSML Controls

Cartesia also supports SSML-like tags inside the transcript. If LyricWinter sees a Cartesia SSML tag in the line, it treats that as deliberate user direction and skips the automatic Cartesia guidance step for that clip.^[2]

• <emotion value="sad"/> directs emotion.
• <speed ratio="1.2"/> changes pacing.
• <volume ratio="0.7"/> changes loudness.
• <break time="1s"/> inserts a fixed pause.
• <spell>ABC123</spell> asks Cartesia to read characters one by one.

Complete Cartesia Emotion List

Cartesia documents primary emotions such as neutral, angry, excited, content, sad, and scared, plus this broader set of supported emotion values.^[1]

A Practical Workflow

• Start automatic: generate the scene normally with Cartesia Sonic 3.5 and let LyricWinter use nearby context.
• Only tag specific misses: if one line needs a clear override, add a Cartesia SSML tag to that line and regenerate the clip.
• Do not mix tag systems: Chatterbox's <sigh> and Fish Audio's [whispering] are not Cartesia controls.
• Prefer natural punctuation first: Cartesia recommends punctuation for normal pauses, reserving <break> for deliberate fixed silence.^[2]

Cartesia Emotion Control FAQ