Practise speaking a foreign language with AI

Why speaking needs its own practice

Most language practice is silent: reading, tapping the right word, listening to a recording. It builds vocabulary and leaves the mouth untrained, which is why people who read a language well often freeze when they have to say something in it.

Speaking improves with speaking, and it improves faster when something reacts to what you actually said rather than to what you meant to say. A teacher or a patient friend is the best version of that. An assistant that listens is the version that is there at eleven at night, does not get bored, and does not mind hearing the same sentence five times.

Check yourself with a translation session

The simplest exercise turns the translation feature around. Instead of translating someone else into your language, you speak the language you are learning and let the app translate you back into yours. If the translation says what you meant, a listener would most likely have understood you. If it does not, you have found the sentence to work on.

1. Tap Translate (the globe). In the sheet titled Translation, turn on "Translate chat and live session".
2. Set From to the language you are learning, rather than Auto Detect, so the app knows which language you mean to be speaking.
3. Set To to your own language.
4. Tap Start on iPhone, or Start Session on iPad and Mac, and say one sentence at a time, pausing after each.
5. Read each pair of bubbles: the first is what it heard you say, the second is its translation into your language.
6. Tap End when you finish, and switch translation off in the same sheet. It is an app-wide setting and stays on for the next session until you do.

Reading the two bubbles

The two bubbles tell you different things. When the first bubble shows a different word from the one you said, the problem is probably pronunciation: that is what a listener might hear too. When the first bubble is right and the translation still misses your meaning, the problem is grammar or the choice of word, and the translation shows what your sentence actually says.

One caution. The model is good at working out meaning from context, much as a friendly listener is, so a sentence with a mistake in it can still come out right. Treat a good translation as "understood", not as "correct". The conversation exercise below is where mistakes get pointed out.

Written sentences can be checked the same way. With translation on, text typed into the box "Ask AI or paste context…" is translated too, which is a quick check on word order and spelling before you try the sentence out loud.

A conversation partner that replies in your target language

For conversation, switch translation off and start a blank session with instructions of your own. The instructions are what turn a general assistant into a patient practice partner, so be specific about the language, the level and what you want corrected. The steps set it up; the examples after them are other instructions worth trying.

1. Open Translate and make sure "Translate chat and live session" is off, so what you say goes into the conversation rather than into a translation.
2. Start a blank session and write instructions such as: "Reply only in Spanish. Keep every answer to one or two short sentences. After each of my turns, correct my mistakes briefly, then carry on the conversation."
3. Set Detail Level to Brief, so each reply is short enough to read quickly and answer.
4. Tell it the situation you want to practise: ordering at a bakery, explaining your job, complaining about a late delivery, small talk with a neighbour.
5. Tap Start and talk. If you want to change what you were saying, start speaking again: the assistant stops its reply when you do.
6. Tap End when you are done. The session stays in Sessions, so the corrections are there to reread tomorrow.

• "Play a shopkeeper in Lyon. Use simple French, and correct only the mistakes that would confuse a listener."
• "Ask me one question at a time about my weekend, in German, and wait for my answer."
• "Reply only in Japanese, at beginner level. If I use an English word, give me the Japanese word for it."
• "Today, correct only my verb endings and ignore everything else."

Read its answers aloud with Karaoke

Karaoke readback is the closest the app comes to a pronunciation drill. It highlights the latest answer word by word as you read it out loud, and the recognition is done on the device by Apple Speech. A word that highlights as you say it was recognised; a word that does not highlight is one a listener might not have caught either.

It listens in the Apple Speech language set in Settings, so set that to the language you are learning first. Then, with the Mic off, tap Karaoke and read the assistant's last reply aloud at a natural pace. Say a stubborn word again until it highlights, then read the whole sentence once more.

Because the text is the assistant's own reply, the practice material is always the conversation you just had: the words you needed, at the level you asked for. For reading a longer text of your own, the guide on rehearsing a talk out loud covers Karaoke in more detail.

A fifteen-minute routine

Short and often beats long and rare. One way to fill fifteen minutes a day:

• Five minutes of conversation in a blank session, with Detail Level on Brief and a new situation each day.
• Five minutes rereading the corrections, then saying the corrected sentences into a translation session to check they now come out as you meant.
• Five minutes of Karaoke on the assistant's last replies, repeating any word that will not highlight.
• Keep the sessions in one project for the language, so the past weeks are in one place rather than mixed with work.

What it is not

It is not a course and not a teacher. There is no lesson plan unless you write one into the instructions, and the corrections are the model's view: usually sound, occasionally wrong or oddly formal. For an exam, a certificate or a job that depends on the language, a qualified teacher who hears you speak is still the right person to judge it.

On privacy: in a cloud session your voice goes from the device directly to OpenAI's Realtime API, and the session is stored on the device; Arte Soft's servers never receive the audio, the transcript or the answers. Karaoke recognition happens on the device. On iPhone and iPad sessions need an internet connection; on a Mac with Apple silicon, local sessions can transcribe, answer and translate as text without the cloud.