It’s easier to hear it once. Below is the same phrase synthesized twice. The recordings differ in only one thing: in the second one, there is a hint in square brackets before the text.
These are not different voices or manual sound editing. Same voice, same text, same “Synthesize” button. One word in brackets was added, and the delivery of the whole phrase changed. The pace, the pressure, the way word endings get swallowed.
Why this wasn’t possible before
Classic synthesis was assembled from studio recordings of a voice actor. A person read text into a microphone for hours, and the main requirement was to read evenly. In a neutral tone, with no spikes, so that any phrase could later be stitched together from these pieces without the joins grating on the ear.
The source material had no emotion by design; it was deliberately cleaned out. That is where the familiar sound of answering machines comes from: technically clean, with nothing human left in it.
The new voices work differently. They grew out of natural speech, the way people actually talk in real life. With pauses, speeding up, tiredness toward evening, a chuckle in the middle of a sentence. In material like this, intonation is always there, and it can’t be separated from the words.
So a new voice doesn’t need to fake emotion. It already has it, and the hint in brackets only picks which one to bring out.
What it looks like in practice
The tag goes right in the text, before the part it applies to. It is not spoken aloud.
Hello! [happily] Your order is on its way, the courier will be there in an hour.
We don’t have a list of allowed words. Almost any word you would use to describe a manner of speaking to a real person will do.
- [happily]
- [sadly]
- [whispering]
- [wearily]
- [sternly]
- [laughing]
- [surprised]
- [calmly]
- [anxiously]
It’s worth experimenting. The model understands descriptions we have never tested, and synthesis costs next to nothing. Trying three versions and picking the best is cheaper than guessing.
To our ears, whispering has the strongest effect. It changes not just the volume but the whole manner of speech.
Where it really helps
An appointment reminder read in a flat voice comes across as a system notification. The same text said calmly and in a human way sounds more like a call from a clinic’s front desk. People listen to the second one to the end more often.
The same goes for good news. “Your order is ready” with a happy intonation and without it feels like two different messages, even though the words are the same.
Personal messages are a separate story. Congratulations, invitations, thank-yous to regular customers. A flat tone works worst here, because a mechanical greeting offends more than no greeting at all.
Where not to use emotions
Pressure through intonation works worst of all. An angry or alarmed voice in a cold call campaign comes across as manipulation. The person doesn’t know your company, and the first thing they hear is pushiness. Campaigns like these collect complaints faster than any others, and on our site you can report a call in a couple of clicks.
Close to this is fake urgency. “Anxiously” in a message about a debt looks too much like a scammer’s trick, and people pick up on it instantly. Along with trust in the message, you lose trust in the sender.
And if the phrase is purely functional, like a time, an address or a verification code, a flat neutral tone is the right one. Joy in a voiced SMS code sounds strange rather than friendly.
Two limitations to know about in advance
Not all voices understand emotions; classic voices can’t do this. In the dashboard, the hint under the list appears only for supported voices, and in the API you can see it in the emotions field of the GET /sounds/voices response. If you add a tag for a voice that doesn’t support it, the tag is removed before synthesis. The text is voiced without it, with no error.
Emotions don’t work in the “Say text” element either. That text is voiced during the call itself, and a classic voice is used there for speed. A new voice would take several seconds to synthesize, and the contact would hear silence. If you need emotion in a scenario, prepare the recording in advance on the “Audio” page and insert it with the “Play audio” element. You’ll also hear the result before your customers do.
Frequently asked questions
How do I add emotion to synthesized speech?
Put a hint in square brackets before the phrase, for example “Hello! [happily] Your order is on its way.” The hint sets how the next part of the text is spoken and is not read aloud. Almost any word works, from “happily” and “wearily” to “whispering” and “laughing.”
Do all voices support emotions?
No, only the new ones. In the dashboard, a hint appears under the list when a supported voice is selected, and in the API you can see it in the emotions field of the GET /sounds/voices response. For classic voices, tags are removed before synthesis. The text is voiced without them, with no error.
Why can’t the old voices do this?
Classic synthesis was built from studio recordings of a voice actor, read in a deliberately flat tone so that phrases could be stitched together. There was simply no emotion in that material. The new voices grew out of natural speech, where intonation is always present, so they don’t need to fake it. You only need to say which one to bring out.
Do emotions work in the “Say text” scenario element?
No. That text is voiced during the call itself, and a classic voice is used for speed. A new voice would take several seconds to synthesize, and the contact would hear silence. Tags are removed there automatically. If you need emotion in a scenario, prepare an audio recording in advance and insert it with the “Play audio” element.
How much does synthesis with emotions cost?
The same as regular synthesis on your plan. An intonation tag counts as part of the text, not as a separate service, and does not affect the price.
You can try it right now. The voice broadcasts page explains how voicing works, and in the dashboard you just open “Audio”, pick a new voice and write text with any tag in brackets. You hear the result right away, before it is sent to anyone.