One voice, 75 languages: what actually changes
Switching locale is more than swapping a voice ID. Numbers, dates, names, and prosody all move, and a few of them break in predictable ways.
A multilingual voice demo usually plays the same sentence in five languages and sounds convincing. Production is harder. The text you send is rarely a clean sentence, and the voice has to survive user names, prices, dates, and abbreviations.
The voice ID is the easy part. The work is in what surrounds it.
Numbers and dates do not translate themselves
The string 10/07/2026 is October 7 in the United States and July 10 in most of Europe. A speech model will read the digits correctly and still say the wrong date, because the ambiguity was in your data, not in the model.
The same applies to currency, decimal separators, phone numbers, and addresses. Normalize these before synthesis, or accept that the audio will be confidently wrong.
Names are the stress test
Product names, people, and places are where accents and fallback rules show up. A voice that handles a paragraph of prose beautifully can still stumble on a brand name with an unusual stress pattern.
A pronunciation dictionary is the unglamorous fix: a small list of terms mapped to a phonetic hint or a respelling. It is also the feature most teams forget to ask about until after launch.
Prosody is locale-specific
Sentence-final pitch, pause length, and the way emphasis lands differ by language. Concatenating a translated sentence onto a voice that was tuned for English produces audio that is intelligible and subtly wrong.
If you are shipping in more than a handful of languages, test with a native speaker on short, real product strings rather than a long paragraph from a demo page. Short strings expose problems faster.
Try it with your own text.
322 voices across 75 languages, plus transcription, on one API key. Free tier included.
Start free