A voice assistant that speaks your language, on the phone itself

Orvena speaks 31 languages, in voices generated on the iPhone, and its interface is in 30. Here is why local matters more for voice than for text, and how it works.

Voice is the most personal data you have

Your voice identifies you the way your face does. Most voice assistants stream audio to a server to be transcribed, which means a conversation has left the room before it has been understood. Orvena's does not. Speech recognition runs on the iPhone through Apple's on-device recognizer, the language model that thinks about your request runs on the iPhone, and the voice you hear is synthesised on the iPhone. Nothing about the conversation depends on a network unless you ask for something that does, such as a web search.

Thirty-one languages

Orvena speaks 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.

It answers in the language you use, message by message. Ask in Danish and it replies in Danish. Switch to German in the next sentence and it follows. You do not configure a language; you speak one.

How a voice is made on a phone

Two speech engines run on the device, each a small neural network. One handles English, Spanish, French, Italian, and Portuguese with a pronunciation dictionary of the most common words in each language. The second handles the other twenty-six, reading text directly. Both are downloaded once: the neural voice is about 330 MB, and the extended voice pack for the additional languages about 400 MB. From then on they work with the radios off.

While you talk, the screen shows what is happening at each moment: listening, thinking, using a tool, speaking. You always know when it is your turn.

The interface too

The app itself is available in thirty of those languages, following the language your iPhone is set to. You can also give Orvena its own language in iOS Settings, independent of the rest of the phone. An Arabic interface is coming.

Limits, honestly

• Orvena reads and writes Norwegian and Chinese, but does not yet have a neural voice for them. Spoken replies in those languages use the iPhone's system voice for now.
• Which languages Orvena can listen to depends on Apple's on-device speech recognition. It covers the languages above on current iPhones, and Orvena picks the recognizer to match the language you are speaking.
• A phone-sized model is smaller than a data-center model, and this shows most in long, specialised conversations. Everyday requests are its strength.
• Voice conversations are part of the one-time Premium Unlock. Typing is free without limits.

Why this matters from Denmark

Orvena is made in Denmark, a country of six million speakers of a language that voice technology has tended to reach late, if at all. Running everything on the phone changes that arithmetic. A Danish voice costs no more to serve than an English one, because the phone does the work. The same is true of Slovak, Estonian, and Vietnamese. That is the practical meaning of thirty-one languages: they are all first-class, because none of them is a line item.