Text to speech pauses and pronunciation: control the voice with text
Most problems with synthetic speech come from the text, not the voice: a pause that is missing, a name read the wrong way, a number in the wrong form. In SuperClonica everything is controlled through the text itself and a few engine options, with no markup language to learn.
This guide shows how the engine reads your text, how to add pauses, how to fix pronunciation and when to change the take or the speed.
How the engine reads your text
The language option must match the language of the text, otherwise the words are read with the wrong sounds.
- Punctuation sets the rhythm: a comma gives a short pause, a full stop a longer one.
- Digits are spelled out as words in 27 languages, but not in Hindi, Greek, Estonian or Croatian, and the words are not inflected.
- Chat shorthand is written out in seven languages: English, Russian, Polish, French, Portuguese, Spanish and German.
- Both can be switched off with --no-numbers and --no-slang.
- SSML tags are not supported: write the text the way it should sound.
Test a tricky sentence free online
On the page of any voice, paste the sentence that gives you trouble, 30 to 100 characters, choose the language and press Voice it. In a few seconds you hear how the engine reads it, and you can change the text and try again. No account is needed.
Pauses
The engine trims silence before and after a phrase; --keep-silence keeps it, which helps when you glue parts together.
- A comma for a breath, a full stop for a clear pause.
- Ellipses for hesitation or suspense: “Well... I am not sure.”
- A new sentence instead of a long one: the voice pauses between sentences.
- For an exact pause, voice the parts separately and insert silence of the length you need in an audio editor.
Pronunciation
Respell what the engine gets wrong, the way it should sound: a foreign name, a brand, a rare word. Write abbreviations either as letters with spaces, A P I, or as the full words. Where grammar matters, write numbers as words, especially in languages where endings change. Split a long compound word with a hyphen if it comes out blurred. Keep the respelled version only in the text you voice, not in the text you publish.
Takes, speed and loudness
Every run is a new take: if a phrase sounds a little off, voice it again, or set --seed with a number to try different takes and repeat the one you like. The default speed is 1.05; a lower value such as 0.95 slows the voice down, a higher one such as 1.1 speeds it up. The --normalize option brings every file to the same loudness, −18 LUFS. All of this is in the free SuperClonica engine: python -m pip install superclonica-cpu, then superclonica --text "Your text" --voice F1 --lang en --seed 7 --speed 0.95 --output take.wav.