Emotional text to speech: how to make an AI voice sound expressive
Flat, even speech is fine for a manual, but a story, a game line or a trailer needs emotion: excitement, anger, calm, fear, sarcasm. Some text-to-speech services offer emotion tags; Supertonic 3, the open model behind SuperClonica, does not. Its emotion comes from three things you control.
This guide explains how to get expressive speech from it honestly: by choosing the right voice, by training a voice from a recording performed with the emotion you need, and by writing the text the way it should be felt. Every voice can be tried for free on your own line.
Where the emotion comes from
There is no slider for happiness or anger, so plan the emotion before you voice the text.
- The voice: each voice keeps the manner of the recording it was made from, calm, lively, gloomy or bouncy.
- The reference: a voice trained on an angry, whispering or cheerful recording speaks every text in that mood.
- The text: punctuation, sentence length and word choice shape intonation and energy.
Choose an expressive voice
Voice descriptions in the catalog say how a voice delivers lines: even, lively, jumpy, restrained, creeping. For energy, pick voices described as lively or energetic, such as energetic narrators and many game and cartoon characters. For calm and warmth, pick even, low voices. For menace, villains with low, measured delivery work best.
Write the emotion into the text
The model reads the whole phrase at once, so punctuation changes the melody of the sentence, not only the pauses.
- Excitement: short sentences and exclamation marks.
- Calm: longer sentences, full stops, no exclamation marks.
- Hesitation or suspense: ellipses and commas for pauses.
- Questions: a question mark raises the end of the phrase.
- Character: the words a person would really say in that mood.
Train a voice with the emotion you need
For a specific mood that no voice has, record a clean reference of 5 to 10 seconds performed in that mood, with its exact text, and train a voice from it in the studio on a computer with an NVIDIA graphics card, or order it on the exchange. Several voices of one speaker, calm, angry and whispering, let you switch emotions line by line.
Test and voice offline
On the page of any voice, type your line of 30 to 100 characters and press Voice it to hear how it handles the emotion. For many lines, download the voice file and use the free SuperClonica engine: python -m pip install superclonica-cpu, then superclonica --text "Get out. Now." --voice villain.json --lang en --output line.wav. It runs offline on an ordinary processor.