Scratch translate block: build a parrot that speaks Japanese
By the end of this evening your child will have a cartoon parrot on screen that takes any sentence they type, turns it into Japanese (or Korean, or Thai), shows it in a speech bubble and says it out loud. The better part comes after, when you make the parrot translate its own Japanese back into English and watch what falls off on the way home.
Translation is the kind of AI most families already use without calling it AI, on holiday menus, on school letters, in group chats with grandparents. A child who has built a translator, and then caught it slipping, will treat the real one differently.
What you need
A computer with internet, and Scratch, the free block-based coding language now run by the Scratch Foundation. Open scratch.mit.edu and click Create to get a blank project.
Two of Scratch’s add-ons do the heavy lifting. According to the Scratch Wiki, the Translate extension uses Google Translate and was made as a collaboration between Scratch and Google, and the Text to Speech extension runs on Amazon Web Services. Both need a connection: the wiki notes the translate block “only works when connected to the internet.”
Build the parrot (four steps)
1. Add the two extensions. At the very bottom left of the editor, under the block categories, there is a purple button. Click it to open the Extensions Library, choose Translate, then click the purple button again and choose Text to Speech. Two new categories appear at the bottom of the block list.
2. Pick a parrot. Click the cat button in the sprite area to open the sprite library, where Parrot sits under Animals.
3. Snap this script together. Each line is one block, stacked in order:
when green flag clicked(from Events)ask [What should I say?] and wait(from Sensing; whatever your child types lands in theanswerblock)set language to (Japanese)(from Text to Speech)say (translate (answer) to (Japanese))(from Looks, with the translate block from Translate dropped into it, and theanswerblock from Sensing dropped into that)speak (translate (answer) to (Japanese))(from Text to Speech, built the same way)
The nesting is the only fiddly part. Build the inside first: drag answer into the white slot of the translate block, set its menu to Japanese, then drag the whole translate block into the say slot.
4. Click the green flag. Type “Good morning, I like noodles” and press Enter. The parrot should show Japanese text and speak it. Try the set voice to block too; the wiki lists five voices, including one called kitten that just says “meow”.
Want Korean or Mandarin instead? Change both menus. The wiki lists 48 languages in the Translate menu, including Japanese, Korean, Chinese (Simplified and Traditional), Indonesian, Thai and Vietnamese, but only 23 for speech. Korean, Japanese and Chinese (Mandarin) can be spoken; Thai, Indonesian and Vietnamese will show in the bubble but not be read aloud. Malay is not in the menu at all, though the wiki describes a workaround: drop a join block containing the word “Malay” into the language slot.
[[photo:1]]
Now make it fail on purpose
Add one more block to the bottom of the script:
say (translate (translate (answer) to (Japanese)) to (English)) for (5) seconds
That sends the sentence to Japanese and straight back. Translators call this a round trip, and it is a cheap way to see what a translation lost without reading the language yourself. Take turns feeding it sentences and score whether the meaning came home intact. Some to try:
- something plain: “The library closes at six.”
- an idiom: “It’s raining cats and dogs.”
- a word with two meanings: “The bat flew out of the cave” against “I left my bat at school.”
- Singlish: “Can or not?”
- a long sentence with three instructions in it.
We did not run every one of these ourselves, and the results change as Google updates its system, so we are not promising which ones break. That is the point of the game. Your child writes down a guess before each test, then checks.
What the parrot is really doing
Google moved Translate to a neural network in 2016, and its own explanation is a good one to read aloud: the new system “translates whole sentences at a time, rather than just piece by piece.” Its researchers reported that this cut translation errors “by more than 55%-85% on several major language pairs,” and the first pair it went live on was Chinese to English.
The same post is unusually honest about the gaps. It says the system “can still make significant errors that a human translator would never make, like dropping words and mistranslating proper names or rare terms, and translating sentences in isolation rather than considering the context of the paragraph or page.” That last bit is why “I left my bat at school” is a good test. The parrot only ever sees one sentence.
It also depends on the language. A study published in 2021 in the Journal of General Internal Medicine asked bilingual volunteers in Los Angeles to check 20 hospital discharge instructions run through Google Translate. Meaning survived in 82.5% of the 400 translations overall, but the spread was wide: 94% for Spanish, 90% for Tagalog, 82.5% for Korean, 81.7% for Chinese, 77.5% for Vietnamese, and 55% for Armenian. The tests were done in 2019 and 2020 with small groups of volunteers per language, so treat the numbers as old and rough. But note who came last. The authors added Armenian and Farsi precisely to test “languages of lesser diffusion”, expecting Google Translate to do worse where it has fewer users, and those two scored lowest.
Japan is living with this right now. Nippon.com reported in July 2026 that Fukuoka’s board of education has given AI voice translation devices to teachers working with children of foreign roots. In the same series, the designer of a school-to-home messaging tool explained why it uses 500 phrases translated by professionals instead of live machine translation: “mistranslations can’t be completely avoided, and there are languages for which machine translation doesn’t work well.”
Where to take it next
Our earlier Scratch chatbot build is the perfect pair: that bot followed fixed rules, this parrot predicts. Our explainer on what a token is picks up from there.
And keep the round-trip habit. Next time a translated school notice or a menu comes back looking a bit odd, send it home again and see what survives. If it matters, ask a person who speaks the language. The parrot is fun. It is not a translator you would trust with the date of an exam.
Was this useful?
Sources
- Translate Extension · Scratch Wiki
- Text to Speech Extension · Scratch Wiki
- Extension · Scratch Wiki
- A Neural Network for Machine Translation, at Production Scale · Google Research
- Found in translation: More accurate, fluent sentences in Google Translate · Google
- A Pragmatic Assessment of Google Translate for Emergency Department Instructions · Journal of General Internal Medicine (via UCLA eScholarship)
- Connecting Foreign Families to Japanese Schools: The E-Tra Note Digital Correspondence Log · Nippon.com
- Learning Japanese Anywhere: Remote Teaching Tech Enables Language Instruction at Multiple Sites · Nippon.com