TTS emotion and tone prompt guide
Default AI voice output is competent but flat. The difference between a robotic read and a believable performance is direction — the same techniques a voice actor gets from a script. Expressive TTS models respond to emotional cues, interjections, and pacing markup, but each engine wants them in a different form. This guide turns the emotion you want into the cues your specific platform understands.
How it works
Pick a target emotion and your TTS platform. Each emotion maps to a set
of reliable techniques: parenthetical direction such as (warmly) or
(laughing), interjection words like “oh” and “hmm” that force a natural
breath, and punctuation pacing — ellipses for hesitation, em dashes for
abrupt stops, exclamation for energy. The tool then formats an example line for
your engine: inline cues for ElevenLabs, a tone instruction for OpenAI TTS, or
SSML <prosody> and <emphasis> tags for generic engines.
Technique reference by platform
ElevenLabs
ElevenLabs responds to inline parenthetical cues placed directly in the text. The model interprets them as stage directions:
(laughing) I can't believe it worked!— triggers a laughing delivery(whispering) Don't tell anyone.— produces a hushed, close-mic quality(sighing) Well... here we go again.— delivers the sigh before the sentence
Punctuation matters too: ellipses slow the pace and imply hesitation; em dashes create an abrupt cut; a question mark at the end of a declarative raises pitch. Combining parentheticals with matching punctuation is more reliable than the cue alone.
OpenAI TTS
OpenAI TTS (tts-1 and tts-1-hd) does not directly support inline cues in the same way. Instead, set a system or instruction-level tone description. For example, passing Speak in a warm, encouraging tone as if coaching a student ahead of the content shapes the overall delivery across the whole request.
Within a request, punctuation and sentence rhythm are the main levers — short, punchy sentences read as energetic; long, flowing ones read as calm or formal.
Generic SSML
For AWS Polly, Google Cloud TTS, and Azure Speech, use proper SSML tags:
<speak>
<prosody rate="slow" pitch="-5%">I have some difficult news.</prosody>
<break time="700ms"/>
<prosody rate="fast" volume="loud">But everything is going to be fine!</prosody>
</speak>SSML gives precise, reliable control but requires more markup overhead. It is the right choice for production audio where consistency matters.
Emotions and their cues
| Emotion | ElevenLabs cue | Punctuation technique | SSML approach |
|---|---|---|---|
| Cheerful | (cheerfully) | ! frequently, short sentences | rate="fast" pitch="+10%" |
| Sad | (sadly) | Ellipses, long sentences | rate="slow" pitch="-10%" |
| Excited | (excitedly) | !, ALL CAPS on key words | rate="fast" volume="loud" |
| Whispering | (whispering) | No !, soft punctuation | volume="soft" rate="slow" |
| Laughing | (laughing) | ”Ha — anyway…” | No direct SSML equivalent |
| Authoritative | (firmly) | Short sentences, full stops | rate="medium" pitch="-5%" |
Tips for expressive delivery
- Layer cues with punctuation. A
(somber)tag plus ellipses and shorter sentences reads as genuine sadness; the tag alone often is not enough. - Keep emotional spans short. Models hold an emotion better over a sentence or two than across a long paragraph — break up monologues.
- Test before committing. Some engines speak the cue text aloud. Generate a few seconds first and fall back to pacing-only techniques if a cue leaks into the audio.
- Match energy to content. Forcing high energy onto somber copy sounds uncanny — let the emotion follow the words, not fight them.