Can one synthetic voice read several languages?

The question comes up the moment a site publishes in more than one language: do you want one voice per language, or a single voice that can read everything? The technical answer tends to surprise people. Yes, one neural voice can in principle read several languages. But "can" is not "should," and the gap between the two is exactly what separates audio people listen to from audio they switch off.

The language is in the request, not in the voice

Modern voice engines are multilingual by design. The reading language is not a property of the voice; it is a parameter of the request. You send text, you say "read this in German," and the engine applies its German rules, whatever voice you picked. So nothing, in theory, stops you from pointing your usual French voice at an English or Spanish text.

We don't do that, and the choice deserves explaining rather than being taken on faith. A French voice reading German reads it with a French accent. That isn't a bad setting, it is what the voice learned: its ear formed on French sounds, and it folds neighbouring sounds back onto the ones it knows. The result is intelligible, but the listener hears the accent before the text, and draws the same conclusion as with a French voice that sounds English: the machine is guessing the language, not speaking it. That is why we pair a native voice with each language and let you choose one per language, rather than stretching a single voice across all of them.

Listing a voice is not reading a language

There is a sharper version of this, and it is the lesson that settled the matter for us. A provider can serve voices labelled in a language without its engine actually being able to read that language. We saw it first-hand: a catalogue sweep had pulled in voices tagged in two South Asian languages, because the label existed. In testing, the synthesis engine refused those language codes: the very same voices read perfectly in a neighbouring language, but not in the one they were filed under. So they were not voices in those languages. They were voices shelved there.

We removed them rather than rig a workaround. We could have transliterated the text into the neighbouring language and had it read that way; to the ear it would almost have worked. But that is selling as one language something that isn't, with the back-up language's accent thrown in. The rule that falls out of it holds well beyond those two cases: listing a voice and being able to make it speak are two different questions, and only the second one counts. It is the same spirit as evaluating a voice yourself instead of trusting a spec sheet: what's printed on the label only binds the person reading it.

What it means for your site

In practice, for a French-language site, you read French with a native French voice: we currently offer twenty-two of them, validated by ear, in different timbres and registers, which is enough to tell an editorial apart from a news brief without changing tools. If your site also publishes in other languages, each one is read by a voice native to it, and you choose which. A translated article keeps its player, its look and its configuration: only the voice changes with the language. That is the whole point of the multilingual setting, and it is also what lets a single site serve two languages with two distinct voices without a second integration.

The number that matters

It is tempting to line up available languages as a selling point. We hold back, for a reason of method. A language is only "switched on" here once its voice line-up has been listened to and validated by a person, never because it shows up in a catalogue sweep. Until a language has its native voices validated, it doesn't appear, even if the provider claims to serve it. So the number that matters isn't how many languages are displayed, it is the certainty that each one that is displayed is actually read, and read well.

If you're torn between a single voice and one voice per language, the test is simple and takes ten seconds: have the tool read a passage in each language on your site, with the voice you intend to use, and listen for the accent. The right setup is the one where you no longer hear where the voice is from, only what it is saying. That's the stance the French text to speech page defends: one language read well beats thirty languages approximated.

Read next

• Add an audio version to your Weebly posts
• Reading a data table aloud: what text-to-speech actually does with it
• Several writers on one account: team, roles and access