How do you practise a language with Advanced Voice Mode?

Speaking is the hardest part of learning a language to practise alone. You can read and do exercises on your own, but conversation needs a partner who answers right away, doesn't lose patience and doesn't mind the same question asked five times. Advanced Voice Mode in ChatGPT comes close to that partner, and since last week it is available to far more people.

OpenAI began rolling it out to ChatGPT Plus and Team subscribers in the ChatGPT app on September 24, over the course of that week, with Enterprise and Edu customers next. When it reaches your account, a pop-up appears beside the voice icon. It is not yet offered in the European Union, the United Kingdom, Switzerland, Iceland, Norway or Liechtenstein.

Why the new voice mode suits language practice

The earlier voice feature passed your speech through three separate steps: one model turned audio into text, GPT-4 answered in text, and a third model read the answer aloud. Along the way, the main model never heard how you sounded. It could not pick up tone, and replies took several seconds on average.

Advanced Voice Mode is built on GPT-4o, a single model trained on text, images and audio together, so your voice goes straight to the model. OpenAI reports that GPT-4o can respond to speech in about a third of a second on average, close to the rhythm of a human conversation. For a learner, that speed matters: long pauses turn a conversation into a series of separate exercises.

With the rollout, OpenAI says the mode understands accents better and runs more smoothly than during its earlier limited alpha test. Two personalization features now work in voice as well: custom instructions, which tell ChatGPT how to respond in every chat, and memory, which lets it recall details you shared earlier. Both are useful for language practice, as you'll see below.

Set up your practice partner

Before you start, decide three things:

1. Your level. Beginner, intermediate or advanced, or a reference such as "about B1." It helps the model choose vocabulary and speed.
2. A scenario. Real situations work better than "let's chat": a bakery, a doctor's appointment, a phone call to a landlord.
3. Your correction rules. How and when you want to be corrected (more on this below).

Put the stable parts in your custom instructions, found in the app's settings, so they apply every time you open voice. Here is an example for a learner of French living in Québec:

I am learning French, intermediate level (about B1), and I live in Québec.
When we speak French in voice mode:
- Speak a little slower than normal, in everyday Québec French.
- Keep your turns short (two or three sentences) so I talk more than you.
- If I don't understand, rephrase with simpler words before switching to English.
- Follow my correction rules: after my turn, if I made an important mistake, say "Petite correction" and repeat my sentence correctly once. Then continue the conversation. Ignore small slips that don't change the meaning.
- Every ten minutes or so, give me one useful expression related to what we're discussing.

A worked example: the bakery

Open the voice screen and start with the situation:

Let's do a role-play in French. You work at a bakery in Montréal and I'm a customer. Start by greeting me.

The exchange might go like this (illustrative):

• ChatGPT: "Bonjour! Qu'est-ce que je vous sers aujourd'hui?"
• You: "Je veux deux croissants et un pain… pain de blé, s'il vous plaît."
• ChatGPT: "Petite correction: Je voudrais deux croissants et un pain de blé. 'Je voudrais' is more polite. Parfait! Voulez-vous autre chose?"

The correction takes a few seconds, then the scene continues. After ten minutes, ask for a quick summary:

Stop the role-play. In English, list the three mistakes I made most often and give me one short exercise for each.

How do you get it to correct you?

Tell it exactly what kind of correction you want. Without rules, you leave that choice to the model, and its idea of helpful may not match yours. Three approaches work well, and you can choose by mood:

• Recasting: it repeats your sentence correctly and keeps going, without explanation. Language teachers call this a recast. It keeps the flow and suits fluency practice.
• Flag and explain: it stops briefly after each important mistake, names it ("that needs the subjunctive") and asks you to repeat the sentence. It is slower but better for accuracy.
• Delayed feedback: it lets you speak freely and lists your mistakes at the end. This works well for confidence and for longer role-plays.

Be specific about what counts as important: verb tenses, gender agreement, a word that changes the meaning. Otherwise, a model asked to "correct everything" will interrupt so often that you stop talking.

Pronunciation is worth testing too. Because GPT-4o hears the audio itself, you can ask, "Did my 'u' in 'tu' sound right?" Treat its answers as a helpful opinion rather than an expert assessment, and check doubtful points with a teacher or a native speaker. If memory is turned on, you can also ask it to remember the mistakes you make most often and to start the next session by practising them.