Here is a scene that plays out in Toastmasters clubs across Malaysia and Singapore every week. A member whose first language is not English writes a speech in careful, slightly formal English, pastes it into a chatbot, and types "improve this." What comes back is grammatically flawless. It is also strangely empty. The sentences are shorter, the words bigger, the phrasing so smooth that it no longer sounds like the person who wrote it. The member now faces a choice: deliver a speech that is technically correct but does not sound like them, or revert to the original with all its small mistakes intact. Neither feels right.
If that is you, this article is written for you. You face a double problem that native speakers mostly do not. The AI can fix your grammar, but it also flattens your voice. The trick is to separate those two jobs completely. Let the model clean up your English, and refuse to let it replace your identity. Here is a workflow for doing exactly that.
You are solving two different problems
A native speaker who asks an AI to polish a speech is mostly asking for style. A second-language speaker is often asking for something more basic and more valuable: correctness. Wrong prepositions, mixed-up verb tenses, a word that means something slightly different from what you intended. These are not creative choices; they are friction, and the audience notices them even when everyone is too polite to say so.
The problem is that a language model does not naturally understand the difference between "fix my mistakes" and "make this sound more professional." It will do both at once, and the second job is the dangerous one. Fixing a wrong tense keeps your message intact. Smoothing every sentence into a single polished register quietly deletes the thing that made the speech yours in the first place.
So the first rule is explicit: never ask an AI to "improve" a speech. Always name the exact job you want done. "Fix the grammar and leave the structure, the word choices, and the phrasing alone" is a completely different instruction from "improve this," and the output is completely different too. This one habit alone will save you from most of the voice-flattening people complain about.
Teach it your vocabulary ceiling
The second thing that goes wrong for second-language speakers is subtler. A model told to polish tends to reach for its favorite words — "leverage," "facilitate," "compelling," "underscore." These are fine words, but they may not be your words. If you would never actually say "I want to leverage this opportunity" out loud, then a speech that tells you to say exactly that has not helped you. It has given you a script you cannot deliver naturally, because the vocabulary is not yours.
The fix is to give the model a ceiling. Tell it, in plain terms, what kind of English you actually speak. For example: "I am comfortable with everyday business English. Do not use words I would need to look up, and do not make the sentences sound like a corporate press release." Or, if you are preparing for a contest where stronger vocabulary is genuinely useful: "You may upgrade two or three words, but list every word you changed and give me a simpler alternative, so I can choose."
The most useful version of this is to have the model give you options rather than decisions. Ask for a "fix grammar only" version and a "fix grammar and slightly raise the vocabulary" version, then compare them side by side. Reading two versions of your own sentence teaches you more about English than any grammar book, because you already know what you meant. The model is simply showing you two ways to say it.
Protect your cultural references
This is the one that matters most for speakers in this part of the world, and it is the one nobody warns you about. When you tell a story that is specific to your life — a hawker stall, a family gathering, a local festival, a habit your parents taught you — a language model does not know what to do with the specifics. Its instinct is to translate them into something universal, because universal is statistically safer. "Eating at the mamak after work" becomes "having dinner with colleagues." "My mother's advice when I left for KL" becomes "advice I received when I moved to the city."
But the specificity is the speech. A story about your mamak, your mother, your town is exactly what makes an audience lean forward. The moment a model replaces your details with generic ones, it has turned a personal story into a template, and no amount of grammatical polish can buy that back.
The rule here is easy to state and easy to enforce: any sentence that carries a proper noun, a specific place, a family detail, or a local reference is yours, and the model is not allowed to touch it. You can even tell it this directly: "Anything that looks like a specific memory or local reference, leave exactly as I wrote it. Only fix grammar around it." Then read the result specifically looking for what got translated. If the mamak became a cafe, change it back and do not apologize for it.
Rehearse the words that trip you
There is one area where AI is useful to a second-language speaker that native speakers rarely think about: pronunciation and delivery. You do not need the model to generate a speech for you to practice. You need it to help you find the places where your mouth is going to fail you on stage.
Before you deliver, paste your final script in and ask: "Read this as a nervous second-language English speaker. Which words or phrases am I most likely to stumble over or mispronounce? Which sentence has a rhythm that is hard to say out loud in one breath?" A model can flag the words with tricky stress patterns — the three-syllable words with unexpected emphasis, the consonant clusters, the idioms that look different from how they sound. It will not pronounce them for you perfectly, but it will tell you where to slow down.
Better still, use a text-to-speech tool to hear the sentence spoken aloud, then shadow it a few times. Say the hard sentence after the voice, then say it again without help. This is rehearsal, not writing, and it is the part of the workflow that actually moves the needle on the night. A speaker who has already said a difficult sentence out loud twenty times does not fear it on stage. A speaker who has only read it silently does.
A small worked example
Here is what the whole thing looks like on a single sentence, using a deliberately generic line so nothing real is at stake. Suppose a member drafts:
"Last year I join the hiking club because I want to become more healthier and also to meet new friend in my city."
Ask the model to "improve" it and you will likely get something like: "Last year I joined the hiking club to embrace a healthier lifestyle and expand my social circle within my community." Correct, but no longer recognizable as the speaker's sentence. "Embrace a healthier lifestyle" and "expand my social circle" are not things most people say about a weekend hike.
Now give the model the right instruction: "Fix the grammar only. Do not change my word choices or sentence structure." You get: "Last year I joined the hiking club because I wanted to become healthier and also to meet new friends in my city." The mistakes are gone — "join" becomes "joined," "more healthier" becomes "healthier," "friend" becomes "friends," "want" becomes "wanted" — and the sentence is still clearly the same person. That is the entire method in miniature. Grammar fixed, voice intact.
Where AI honestly cannot help you
Let me be clear about the limits, because the worst version of this advice is the promise that a chatbot will turn a hesitant speaker into a confident one. It will not. A language model cannot teach you the courage to walk to the front of the room. It cannot fix the dry mouth, the racing heart, or the fear of forgetting your own name. It cannot replace the club itself — the real audience, the evaluator who looks you in the eye, the applause that tells you a line landed.
What it can do, reliably and cheaply, is the mechanical work that second-language speakers spend the most anxious hours on: checking grammar, checking that a word means what you think it means, flagging the sentences that are too long to say in one breath, and giving you a low-stakes reader that will not flatter you the way a kind friend will. Used that way, it does not write your speech. It hands you back a cleaner version of the speech you already wrote, so you can spend your energy on the part only you can do — standing up and saying it.
The workflow, end to end
Putting it all together, here is the order I actually recommend for a second-language speaker preparing a speech:
- Draft in your own words. Write the speech the way you would say it, mistakes and all. Do not ask for help yet.
- Fix grammar only. Paste it and ask explicitly: fix grammar and nothing else. Do not let it restructure or reword.
- Set a vocabulary ceiling. Tell the model what kind of English you speak, and ask for a side-by-side "slightly upgraded" version only if you want one. Choose, do not accept by default.
- Protect your references. Read the result hunting for any specific detail that got translated into something generic, and restore every one of them.
- Flag the trip hazards. Ask which words and sentences a nervous speaker is likely to stumble over, then practice them aloud with a text-to-speech tool.
- Deliver it standing up. None of the above matters if you do not actually say the speech out loud, at least once, on your feet.
The last step is where the speech becomes yours again, no matter how much help you had on the way there. The AI can correct your English, but only your voice — your memories, your jokes, your particular way of landing a line — can make the room feel something. Keep that division clean, and the model stays a useful editor instead of becoming a ghostwriter you never asked for.
So the next time you sit down with a draft, try the two-step version first. Ask it to fix your grammar and nothing else, then read the result and notice how different it feels from the "improve this" output you are used to. You will probably find that most of your English was fine all along — it just needed a little cleaning up, not a new voice.
