There’s a completely normal way people in Taiwan talk: mixing Chinese and English in the same sentence. “This meeting 幫我 booking 一下”—“book this meeting for me.” We write the same way. A paragraph of Chinese with a few English words dropped in reads perfectly fine to us.
Hand that same script to an AI voice and it can turn into a car crash in one second. A model number spelled out letter by letter, “APP” read as “A-P-P,” a brand name mangled beyond recognition. One moment it sounds professional, the next it’s completely off.
So I took my real work scripts and tested code-switching systematically. This is the three things that break most often, my pre-processing routine, and a multilingual template you can copy directly.
The short version: it’s not the AI, it’s the script
Most failures aren’t the model reading wrong—it’s the script being unclear. Where Chinese ends and English begins, how a number should be read, which pronunciation a proper noun needs—if you don’t give that information, the AI guesses, and guessing wrong is normal.
So this piece isn’t about “which tool handles multilingual best.” It’s about how to write a script that any tool can read correctly. Get the script right and switching tools won’t break it.
The crash scene: three real failures
Three real ones I hit, so you can see how everyday this is.
First, a phone model. The script said “iPhone 15 Pro.” The AI once read it as “I-P-H-O-N-E one-five Pro”—spelling iPhone as letters and reading 15 as “one five.” An entire product video, instantly unprofessional.
Second, the word APP. “Open our APP.” Some tools read it “A-P-P,” some read it “app.” Both are defensible, but inconsistent within a single video sounds sloppy.
Third, a brand name. An English brand name read with a Chinese-pinyin accent, like a foreigner seeing the word for the first time, stress all wrong.
The common thread: these are all boundary problems. The AI isn’t sure if this is English or Chinese, whether to spell it or say it, where the stress goes.
My pre-processing routine: tidy the script first
I built a five-minute “tidying” habit before every recording, and the failure rate dropped by more than half.
First, put a space between languages. “這個meeting” versus “這個 meeting”—the AI reads the latter far more accurately. One space tells it “the language switches here.”
Second, pick one spelling and stick to it. The same word, the same form, for the whole video. APP in caps everywhere, email in lowercase everywhere—not app here, APP there, App elsewhere.
Third, mark out anything that should be spelled. If you want “A-P-P,” write “A P P” with spaces in the script. If you want it read as “app,” write “app.” How you want it said is how you write it.
Three small steps, but once they’re habit, you’ll dodge the landmines while writing in the first place.
Numbers, dates, percentages: the three worst offenders
Numbers are the disaster zone in mixed-language scripts because they test two things at once: Chinese or English, and which reading to use.
Whole numbers and decimals. “3.5 萬”—should it be “three point five wan” or “three five wan”? If you don’t specify, it might try both. My rule: where a number must be precise, write it out in Chinese characters: 三點五萬.
Percentages. “15%”—some tools say “百分之十五,” some say “十五趴.” Writing it as 百分之十五 is the safest.
Dates and years. “2026”—“二零二六” or “兩千零二十六”? For years, 二零二六 usually sounds more natural. For “the 1990s,” write 一九九零年代. I hard-code all of these in Chinese.
My iron rule: if a number must be exact, write it in Chinese. If it should sound casual, pick one reading and use it for the whole video.
Brand names and proper nouns: build a pronunciation table
The best investment I made: a personal “pronunciation table.”
Next to every brand, product, or person that comes up often, write how you want it read. For example:
- APP → write “A P P” to spell it, or “app” to read it (pick one)
- iPhone 15 Pro → write “iPhone fifteen Pro,” spelling the number in English
- A foreign name → add a phonetic hint in Chinese “（唸：史蒂夫）”
The table grows over time, but you build it once and save yourself re-testing on every video. Mine lives in a note I glance at before each recording.
True multilingual scripts: Chinese, English, and Japanese
Some scripts aren’t just Chinese-English—they mix in Japanese or Korean. I tested a Chinese-English-Japanese script, and the result was better than expected, as long as you put spaces at each language boundary.
But there’s one trap: three languages in a single sentence tends to scramble everything. Chinese-English is fine, Chinese-Japanese is fine, but “Chinese with English and Japanese mixed in” makes some tools start switching channels. My fix: separate paragraphs by language, and keep a single sentence to at most two languages.
How to judge a tool’s multilingual skill
If you want to test multilingual ability before paying, here are three tests:
First, a sentence of everyday code-switching. Something like “我等等要 join 一個 call，你幫我 book 一下餐廳.” Listen to whether the switch is smooth and whether the English words carry a natural accent.
Second, an English brand name inside Chinese. Check whether the English is read with an English accent or a Chinese one. A good tool reads the English word with English phonetics, not Chinese syllables forced onto it.
Third, a string of numbers and percentages. This is the mirror test—it reveals the tool’s grasp of number rules in one shot.
A multilingual template you can copy
Here’s one I actually use. Swap in your own words and it works:
「大家好，歡迎來到今天的分享。這個 app 呢，我們從頭到尾改了一遍。打開之後，你會看到一個全新的 dashboard，速度比上一版快了百分之三十。我們這次特別針對 iPhone 十五 Pro 做了優化，拍照的部分，進步最有感。有興趣的話，歡迎到官網下載，連結我放在下面。」
Notice the details: spaces between languages, 百分之三十 written in Chinese, iPhone 十五 Pro with the number spelled out in Chinese. With those in place, any tool will read it cleanly.
One caution: don’t over-tidy
Tidying matters, but you can take it too far. I once chased perfection and rewrote every English word into Chinese. The result lost the original flavor and sounded odd.
The rule is: keep the English where it belongs, and just make the boundary clear enough for the AI to judge. Code-switching is the rhythm Taiwanese readers actually know—forcing it into pure Chinese sounds more awkward, not less.
Also worth noting: some tools let you set a “default language” for a voice. If a voice is built for Chinese, its English will carry a slight Chinese accent no matter what you do. If your script leans English, pick a voice built for English instead. Matching the voice to the dominant language matters more than any amount of script tidying.
Final thought
Code-switching is how Taiwanese creators actually talk—it’s a feature, not a flaw. You don’t need to rewrite your script into pure Chinese to please a tool. Spend five minutes tidying, and the AI will read your copy smoothly and naturally.
Multilingual is just one part of voiceover. For the fundamentals, read What Is TTS. If you make short videos, these eight tips are just as useful.
