An AI voice assistant that speaks your language: 31 of them, on the iPhone

Orvena is an AI voice assistant that speaks 31 languages with voices generated on the iPhone, and its interface is in 30. Here is the list, how to set it up, and the details worth knowing for your language.

AI voice assistant languages: the full list of 31

English (American and British voices), Spanish, French, Italian and Portuguese have dedicated voices with their own pronunciation dictionaries. The other 26 share one engine and a set of ten voice styles: Arabic, Bulgarian, Croatian, Czech, Danish, Dutch, Estonian, Finnish, German, Greek, Hindi, Hungarian, Indonesian, Japanese, Korean, Latvian, Lithuanian, Polish, Romanian, Russian, Slovak, Slovenian, Swedish, Turkish, Ukrainian and Vietnamese.

You do not set an answer language. Orvena replies in the language you wrote or spoke in, message by message, so you can switch mid-conversation and it follows.

That holds for Orvena's own model. Apple Intelligence, which Orvena can also answer with on iOS 27, covers fewer languages. Apple lists these languages for Apple Intelligence: English, Danish, Dutch, French, German, Italian, Norwegian, Portuguese, Spanish, Swedish, Turkish, Vietnamese, Chinese (simplified and traditional), Japanese and Korean. Orvena asks iOS whether Apple's model supports your iPhone's language, and in a language it does not support, it shows "Apple Intelligence doesn't support this language yet. Download Orvena's own model in Models to chat in it." The default engine setting, Automatic, only picks Apple Intelligence when it supports your iPhone's language.

Setting the spoken language

By default the recognizer is set to Automatic: Orvena listens in the language of the current conversation, and if there is none yet, in your phone's language. If you always speak one language, pin it under Settings, then Voice, and the recognizer stops guessing. The first time a language is used, iOS downloads its on-device recognition model, so the "Getting ready…" screen before you speak stays up a little longer that time. iOS only keeps a handful of recognition languages installed at once; Orvena manages this and always keeps your phone's language and the one you pinned.

Choosing a voice

English has eight voices. Heart is the default; the other American voices are Bella, Nicole, who whispers, Michael and Fenrir, and the British ones are Emma, George and Fable. Spanish, French, Italian and Portuguese each have a voice of their own. For the 26 other languages you choose one of ten styles, Female 1 to 5 and Male 1 to 5, and the style carries across languages. Each English voice and each of the ten styles has a preview button in Settings; the Spanish, French, Italian and Portuguese voices are not picked but switch in on their own when a reply is in that language.

The downloads

Voice comes in two packs, downloaded after you unlock it and finished in the background: about 345 MB for English, Spanish, French, Italian and Portuguese, and 400 MB for the other 26 languages. If you only need one of them, the other can be removed in Settings. Both come from Orvena's mirror with every file checked against a pinned hash, the same way the language model does.

How the voices are made

The first engine is Kokoro, an 82-million-parameter model that works from pronunciations rather than spelling. For English it carries pronunciation dictionaries of more than 180,000 words inside the app and a small neural network for anything not in them. For Spanish, French, Italian and Portuguese it downloads a pronunciation dictionary per language, sounds out rare words from letter rules, and spells numbers the way a speaker of that language would. The second engine, Supertonic, reads text directly without a pronunciation step, which is what lets it cover 26 languages in one 400 MB download. Both pause about a quarter of a second after a sentence, and Supertonic also pauses half that at a comma. Both read symbols aloud: percentages, currencies and temperatures are spoken as words in English, Spanish, French, Italian and Portuguese.

Notes per language

• Norwegian. Orvena reads and writes it, and speaks it with the Danish voice, since written Norwegian and Danish are close enough for the Danish voice to read it well.
• Malay and Galician. Spoken with the Indonesian and Portuguese voices respectively, for the same reason.
• Chinese. Understood and written, but there is no neural voice yet, so spoken replies use the iPhone's built-in voice for now.
• Arabic. The voice is there, but the interface is not translated into Arabic.
• Japanese and Korean. Spoken in shorter phrases internally, which suits both languages; you will not notice it except that pauses fall naturally.
• Siri. The Siri phrases for Ask Orvena and Talk to Orvena are translated into the 29 interface languages besides English, for example "Frag Orvena" in German, "Spørg Orvena" in Danish and "Demande à Orvena" in French. Whether Siri responds to them depends on the languages Siri supports.
• Any language. Whether Orvena can listen in it depends on Apple's on-device recognizer. The picker only offers languages it can both hear and speak.

Common questions

Which languages does Orvena speak?

Thirty-one: English, Spanish, French, Italian, Portuguese, German, Dutch, Danish, Swedish, Finnish, Polish, Czech, Slovak, Slovenian, Croatian, Hungarian, Romanian, Bulgarian, Greek, Ukrainian, Russian, Turkish, Estonian, Latvian, Lithuanian, Arabic, Hindi, Indonesian, Vietnamese, Korean and Japanese. The interface is available in 30 languages.

Are the voices generated on the phone?

Yes. Both speech engines run on the iPhone, and with Orvena's own model or Apple Intelligence the reply is written there too. If you choose Private Cloud Compute or a cloud model, the reply is written on that provider's servers, and a live voice session on OpenAI's realtime model with your own key streams your audio to OpenAI. Each of those asks for consent first.

Can it translate?

You can ask it to translate a sentence or a message, and it will. Otherwise it simply answers in the language you used.

Is the interface in my language?

In 30 languages, listed on the App Store page. Arabic is spoken but the interface is not yet translated.