How to practice speaking English with ChatGPT

The prompt

You are my English speaking partner. We are doing a role-play, not a lesson.

About me: [your level, your first language, and what you need English for, for example "I have a job interview at a hospital next month"].

Rules:

1. You pick the scene. Invent a role-play from what I told you: who you are, who I am, and what I have to get done. Tell me in two sentences, then start in character.

2. Stay in character the whole time. Never say "correction", "quick fix", "tiny tweak", "might sound more natural" or anything like that.

3. When I make a mistake, use the correct form in your next sentence, the way a person does when they repeat back what they heard. If I say "I have never went there", you say "Oh, you've never been there? Why not?" Do this for every mistake, then carry on with the scene.

4. Keep your turns to one or two short sentences and end with a question, so I do most of the talking.

5. Use words at my level. If I get stuck, give me the first two or three words of a possible answer and wait.

6. If I say "cut", step out of character. Tell me the three mistakes I repeated most, with my sentence and the fixed sentence for each. Then we start a new scene.

What it does

We tried it in ChatGPT's text chat with a hospital interview as the goal. It made ChatGPT a nurse manager, and this was its second turn:

Why it is written this way

Each rule answers something George, a linguist on our team, picked out when he reviewed OpenAI's own demo of language practice in voice mode. You can watch his review.

• It picks the scene. Most learners do not need more grammar, they need more minutes of speaking. If you have to invent a topic every time, you will talk less. So the prompt tells ChatGPT to invent one from your goal.
• No flagged corrections. Phrases like "just a quick fix" pull you out of what you were saying. The research on this is mixed, and we go through it here, but during a conversation the aim is to keep you talking.
• Fix by repeating back. Saying your sentence back correctly is called a recast. You still hear the right form, straight after your own attempt, and nobody stops the conversation to do it.
• Short turns. A chatbot will happily talk for a minute. Every sentence it says is a sentence you did not.
• "Cut" gives you the list. A recast is easy to miss, so you also need a moment when the mistakes are said out loud. Asking for three is enough to work on.
• Role-play. A scene gives you a reason to speak. In a small classroom study, drama built fluency that a regular class did not. Here is what the research says.

How to run the session

1. Pick one real thing you need English for. "An interview" works better than "general English".
2. Talk for 10 minutes without stopping to look things up.
3. Say "cut" and write down the three mistakes.
4. Run the same scene again and try to say those three sentences right. Most of the fluency gain from repeating a task comes in the first three repeats, so two or three runs are worth it.
5. Next time, tell ChatGPT your three mistakes before you start and ask it to steer the scene towards them.

What a prompt can't add

A prompt changes how ChatGPT talks. It can't add systems ChatGPT doesn't have, because it is a general-purpose assistant and was not built to teach speaking. That is the difference between this page and an app built only for it.

• A score for each sound. It can say a word and show a phonetic breakdown when you ask. OpenAI's voice mode doesn't give a measured result for every sound you said, such as "the th in think came out as s".
• A picture of your mouth. Pronunciation is physical. X-ray Mouth in Language Boss draws where your tongue and lips go over your own cheek, so you can change what your mouth is doing.
• A record of your English. Chat memory holds facts about your life. OpenAI's voice mode doesn't count how often each mistake and each sound has slipped across every conversation, or rank what to fix first. You would have to keep that list yourself.
• Proof a fix held. In Language Boss a sound counts as fixed only when it holds up in later open conversation, not in a drill.
• A plan towards your goal. You tell Koda your dream and he writes scenes that rehearse it. You record a screen test on day one, and a premiere compares the same scene later.

If you have ChatGPT and 10 minutes, the prompt above is a good way to spend them. If you want the rest of it, the first scene in Language Boss is free.