Can ChatGPT or Gemini Check Your English Pronunciation?

What voice assistants can and cannot tell you about your English pronunciation, a two-minute test to see which one you have, and what a real pronunciation score returns instead.

Published on September 8, 2026

Direct answer

ChatGPT can describe how a word should sound and can react to what it heard, but it cannot tell you reliably which word you said wrong, because nothing in it produces a measurement of your audio against the expected sounds. Gemini works the same way for this purpose: neither product exposes a score per sound, so the verdict you get back is generated in conversation rather than computed from your audio.

That is not a reason to stop using them. It is a reason to use them for the half of the job they are good at, and to use something that measures for the other half.

What a voice assistant is doing when you speak to it

Voice modes in ChatGPT and Gemini are built for conversation. OpenAI describes its realtime audio stack as a way to “build speech-to-speech agents that listen, reason, speak, and call tools” — the goal is an agent that answers you, not an instrument that grades you.

Two consequences follow directly from that design.

First, recognition is built to succeed. It is supposed to understand you through background noise, a cold, a fast sentence and an accent. The errors you want flagged are exactly the ones it was engineered to absorb.

Second, the reply is generated. When you ask “how was my pronunciation?”, the model writes an answer that fits the conversation. Ask twice and you can get two different answers about the same audio, because nothing was measured in between.

What they genuinely help with

It is good for the parts that need a partner rather than a measurement: rehearsing an answer, being asked follow-up questions, and hearing an explanation of where the stress goes in a word. That is real value — the value of a study partner who talks back — and it covers most of what people actually ask a chatbot to do.

What they cannot do

Name the word that failed. A general comment about your speech is not the same as “the second syllable of comfortable was the one the listener lost”.

Give you the same verdict twice. Feedback that changes between attempts cannot tell you whether this attempt was better than yesterday’s.

Disagree with you. A model tuned to be helpful is a bad judge of whether you were understood. The model is built to understand you despite an accent and to be agreeable when it answers, so the two things it does best are exactly the two things that hide your errors.

This is the same trap that catches most language apps, for the same reason: why your English app does not correct pronunciation covers the recognition-versus-assessment problem in full.

The two-minute test

Run this before you trust any AI feedback on your speech.

1. Say the same sentence twice. Record or speak one sentence, ask for feedback, then repeat the identical sentence in a new chat and ask again. A measurement gives you the same verdict twice; an impression does not.
2. Break a word on purpose. Choose a word you can say correctly, deliberately break one sound in it, and see whether the assistant names that word or carries on with the conversation.
3. Ask which word failed, not how you did. Ask for the specific word and the specific syllable rather than a general assessment, and note whether the answer points at anything you can act on.

If step 1 gives you two different verdicts and step 2 goes unnoticed, you have a conversation partner. Use it as one.

What a pronunciation score returns instead

Pronunciation assessment is a separate technology with a separate output. Microsoft’s speech service, for example, lets you request results at phoneme, word or full-text granularity, and its documentation states that “syllable, word, and full text accuracy scores are aggregated from the phoneme-level accuracy score”. See the score list and its granularity settings.

The difference that matters to you is not the jargon. It is this: a pronunciation assessment engine returns a number per sound and per word for the same audio every time, which is why it can point at the syllable that failed instead of describing your speech in general. The underlying method has a name and a long research history — what GOP scoring is explains how a score per sound is actually computed.

Google’s real pronunciation tool is not Gemini

Worth knowing if you searched for “Gemini pronunciation”: Google shipped speaking practice inside Search, not inside its chatbot. Its research team describes the feature as short sessions where learners answer real-life prompts using given vocabulary and receive semantic and grammar feedback, launched for Android users in Argentina, Colombia, India, Indonesia, Mexico and Venezuela.

Google has a real speaking tool, but it is not Gemini: Google put speaking practice inside Search itself, and that feature was built with its own analysis of your spoken answer rather than a chat model improvising a reply. Its feedback is still aimed at whether your answer worked as language rather than at which sound broke, and it is not available everywhere.

How to use both

Use the chatbot for the things that need a partner and something that measures for the things that need evidence.

Phonema does the measuring half: it scores each word you say against the expected sounds and shows you the one that broke, on your device, with no recording leaving the phone. For the other checks that cost nothing, six ways to check your English pronunciation lists what each one can and cannot see.

Bottom line

Ask ChatGPT or Gemini to explain a word, to model it, and to talk with you — they are good at that and free at the entry level. Do not ask them whether you said it right. Run the two-minute test once and you will see why: the answer changes, and an answer that changes cannot show you progress.

Related sounds

• θ — as in think
• ɹ — as in red
• ɪ — as in sit
• i — as in see