Why it happens
- Only the text was provided, with no listener, subtext or current state.
- Lines are too long and too written to sound like speech.
- Emotion is described vaguely, just “sad” or “angry”.
How to fix it
Tell the voice model who the line is for, why it's said and what the speaker is hiding.
Break long sentences into short ones, drop written-style connectors and mark pauses.
Describe emotion with intensity and delivery: lowered voice, trembling ending, stopping mid-sentence.
Exclude explicitly: no announcer tone, no over-acting, don't change the perceived age.
If it's overdone, lower the intensity and shorten the description rather than piling on adjectives.
Before and after
Fails
Character: Lin Xia; emotion: scared; line: “The recipient of this order is actually my sister who has been missing for a long time, how could that be possible?”
Works
Character: 25-year-old night-shift clerk, alone in the store; talking to: herself; subtext: can't believe it, afraid it's true; line: “That's impossible…”; delivery: low voice, holds her breath after; no announcer tone, no over-acting.
Prevent it next time：Write dialogue as short lines in the script, and always include listener and subtext fields in voice prompts.
Worked examples