For language learners, the hardest skill to acquire is speaking without embarrassment. AI now offers a tireless conversation partner available around the clock, and two major players—ChatGPT’s voice mode and Duolingo—are betting heavily on it. But while these tools can dramatically increase speaking volume, they are not a substitute for a human teacher. Understanding their strengths and limits is essential to using them well.
- ChatGPT Voice: Offers three modes—Live, Advanced, and Standard—with plan-based limits; Live allows natural interruptions and simultaneous listening and speaking.
- Duolingo’s AI push: Video Call and Roleplay are moving from the premium Max tier to Super, with free users gaining more speaking exercises and new Speaking Adventures lessons.
- Three key risks: AI can produce inaccurate corrections, fail to catch pronunciation errors, and encourage dependency by doing the learner’s work.
ChatGPT Voice: A Practice Partner You Must Configure
OpenAI’s help documentation describes three voice modes. Live listens and speaks simultaneously, enabling interruptions and more natural conversation. Advanced is the earlier real-time experience. Standard works in turns, transcribing your voice before generating a response. In Settings > Voice > Language, you can set your primary language for better comprehension, and you can ask the assistant to switch languages mid-conversation.
Limits depend on your plan. As of November 2026, Live mode in Chat is measured in 24-hour periods: the free plan has limited access to GPT-Live-1.1 Quip, Go provides 3 hours with that same model, and Plus provides 3 hours with GPT-Live-1.1.
The key is your opening instruction. Effective prompts include: “Let’s practice English at B1 level. Speak slowly, use short sentences, and ask me one question at a time”; “Pretend I’m a customer in a London shop and you’re the clerk. Stay in character”; and “Don’t correct me while I speak. At the end, give me a list of my three most frequent mistakes and how to say them correctly.” That last request matters: correcting every sentence breaks the flow, while correcting nothing lets errors persist. To avoid reconfiguring daily, save these instructions in custom settings.
On privacy, OpenAI states that audio clips from Live and Advanced modes are stored with the chat transcript for 30 days and are not used for training unless you enable recording-sharing options in Settings > Data Controls.
Duolingo: From Premium Extra to Core Practice
Duolingo has shifted AI from a paid add-on to part of everyone’s practice, though video-call conversation remains a paid feature. Its 2026 shareholder letters detail the strategy.
Duolingo Max launched as the tier above Super Duolingo. Its March 2023 presentation included Video Call, real-time conversations with the character Lily that remember your previous session and provide a transcript for review, and Roleplay, guided scenarios like ordering coffee in Paris with feedback on accuracy and complexity.
In its February 26, 2026 shareholder letter, the company announced it would move Video Call to the Super plan and test changes to Max, arguing this would bring conversation practice to roughly ten times more learners. It also announced that all users, free and paid, could answer most exercises by speaking, plus a new free lesson type called Speaking Adventures.
The May 4, 2026 first-quarter letter confirms the Video Call expansion to Super was beginning, not complete. For free and paid users, it mentions three additions: answering almost any exercise by speaking instead of tapping words, cards for saying words and phrases aloud, and Speaking Adventures, real-world tasks solved by talking to characters. Duolingo also claims the average number of words spoken per user in Video Call has more than doubled over the past year.
These letters do not detail availability by country or course; the 2023 presentation limited Video Call and Roleplay to specific combinations, all from English. Before paying, check which features appear for your course and plan in the app.
Where AI Gets Language Learning Wrong
Three fronts: it can correct you incorrectly, it may not perceive how you pronounce things, and it can reduce your effort. None invalidates the tool, but all three change how you should use it.
Errors. OpenAI’s European terms of use note that outputs may not always be accurate and should not be used as a sole source of truth. In languages, that means grammatical explanations that sound confident but are wrong, or expressions nobody uses. When a correction surprises you, verify it with a dictionary or grammar reference. Duolingo’s own blog acknowledges the technology is not perfect and invites users to report errors by long-pressing the message.
Accent. ChatGPT Voice’s help warns that transcripts are not verbatim and may not match what you said, especially with background noise or fast conversation. For learners, this has an uncomfortable effect: if the system interprets what you meant to say, it may accept a mispronounced word without warning. A chatbot understanding you does not prove a native speaker would. For fine pronunciation, a human remains the better judge.
Dependency. The risk is that AI does the work meant for you: translating every sentence before you say it or filling in what you forget. It helps to ask it not to give you the word until you have tried, and to distrust easy praise, because chatbots tend to agree with you. The same study principles apply as for any subject: use AI to support your effort, not replace it.
And a useful distinction: practicing is not translating. If you need to understand a text or write an email in another language, other tools are better suited.
Duolingo’s move is the best signal of where this is heading: a company that made money selling AI conversation as a premium extra is now distributing it to far more users because, by its own account, conversation practice is a critical learning feature. The diagnosis is right. What remains unproven is whether talking more with a machine translates into talking better with people—and none of these companies has demonstrated that with public data yet. For now, the sensible combination is AI for daily volume and a human, even occasionally, to tell you the truth about your accent.