Text to speech in 31 languages: one voice for every language

How one voice speaks 31 languages

A voice in Supertonic 3 describes how someone sounds, not which language they speak. The language is chosen for each text with a language code, so the same voice file works for all 31 languages. The accent of the original recording can still be heard a little, which suits characters and is barely noticeable in narrators.

Try it on the page of any voice: pick the language next to the text field, type a phrase of 30 to 100 characters and press Voice it. On your computer, the SuperClonica engine takes the code with --lang, for example superclonica --text "Hola, ¿qué tal?" --voice voice.json --lang es --output hola.wav.

English text to speech

Code en. English is the most common language of the catalog voices: narrators, game and film characters, announcers. Numbers, dates and units are read as words; write abbreviations the way they should be pronounced.

Spanish text to speech

Code es. Write accents and the opening ¿ and ¡ marks: they help the model with stress and intonation. Every voice can read Spanish, including characters recorded in English.

Portuguese text to speech

Code pt. Keep the diacritics (ã, ç, é) in the text. The model reads one Portuguese; Brazilian or European wording is up to your text.

German text to speech

Code de. Long compound words are read correctly as they are; write umlauts and ß, not ae or ss, for the best pronunciation.

French text to speech

Code fr. Keep accents and the apostrophes of elision (l’, d’, j’): they change how words are linked when spoken.

Russian and Ukrainian text to speech

Codes ru and uk. Write in Cyrillic. In Russian, write ё where it belongs; stress marks are not needed. Numbers are read as words in both languages.

Japanese text to speech

Code ja. Write in the usual mix of kanji and kana; for an unusual reading of a name, write it in kana. Japanese suits anime and game characters: many of them can say a line in the original language.

Korean text to speech

Code ko. Write in Hangul with normal spacing between words; the spacing helps the model with rhythm.

Hindi text to speech

Code hi. Write in Devanagari script: the model is made for it, while Hindi written in Latin letters may be read less accurately. Every voice in the catalog can read Hindi.

Arabic text to speech

Code ar. Write in Arabic script; vowel marks on ambiguous words may help the model choose the right reading.

Turkish, Indonesian and Vietnamese text to speech

Codes tr, id and vi. Keep the special letters and tone marks of each language (ç, ğ, ş, ı in Turkish; the tone marks in Vietnamese): without them words are read differently.

Italian, Polish and other European languages

Codes it, pl, nl, sv, da, fi, cs, sk, sl, hr, hu, ro, bg, el, et, lv and lt. Write each language with its own letters and diacritics; Greek and Bulgarian in their own alphabets.

Numbers and dates in every language

The SuperClonica engine spells numbers out as words in 27 of the 31 languages, so 2026 or 150 is read naturally. Grammatical cases are not added: where a language needs a special form, such as an ordinal, write it in words. The engine option numbers=False passes digits to the model as they are.