krnna
Vowels are dropped. This is karanna, to do, with the vowels left out, and readers fill them in without noticing.
Singlish research
Type Singlish, get English.
Research by Tishan Yasiru on reading Sinhala the way Sri Lanka actually types it.
This page saves what you type and the English it returns, to study how people really write Singlish and to improve the translator. No name, email, IP address or device details are saved. By translating, you agree to this. Details are under What we save.
You typed
English
What you type here is saved for the research. See What we save, below.
Usage
Last seven days
The language
Translation systems and keyboard converters expect clean Sinhala script or careful spelling. Recent public benchmarks show that even large general language models handle Sinhala typed in English letters far worse than Sinhala script, and that bigger models do not close the gap.
krnna
Vowels are dropped. This is karanna, to do, with the vowels left out, and readers fill them in without noticing.
mt
One skeleton, several words. In one sentence it reads as to me, in another as something else. Only the rest of the message decides.
meeting ekata
English words carry Sinhala grammar. The ending turns meeting into to the meeting.
one
A false friend. In Singlish this is a Sinhala word meaning need, not the number.
hri, hari, harii
No fixed spelling. Three people write the same word three ways, and one person changes between messages.
The research
The problem
Real chat is noisy by nature. Vowels go missing, words get shortened, English slips in mid-sentence, and spelling follows the typist, not a dictionary. Systems trained on clean script never see any of this, so they fail on exactly the messages people actually send.
The data
A collection of everyday Singlish conversation, written the way people type it, with the noise left in on purpose. Each message is paired with its meaning in Sinhala script and in English, so one collection can teach two directions: Singlish to Sinhala, and Singlish straight to English. Every batch passes checks for duplicates, contamination and invented content, and a native speaker reviews it.
The system
A translation system trained on that collection. It reads the whole message rather than matching word by word, which is how it settles a skeleton like mt from context. It reads text as raw characters instead of a fixed vocabulary of word pieces, so an unusual spelling is a small difference rather than an unknown word.
The measurement
Cases the system has effectively seen before and cases it has not are scored separately, and the hard cases are reported on their own instead of disappearing into an average. A version that looks better overall but worse on the hard cases does not ship.
The goal
A version small enough to run on a phone or a laptop, so a message never has to leave the device to be understood. What you are using on this page is the current version running on a server. Shrinking it to run on the device is the next step, and the messages people type here are what will show whether it is ready.
What is not said here
The model behind this page has a name, a size, a training story and scores. None of them are published here while the work is under review.
How it works
Limitations
It makes mistakes, and a confident mistake can read fluently.
It works best with everyday chat. Formal or technical text is outside what it has seen.
Singlish spelling varies a lot, so the same message spelled differently can get a different translation.
Output is English only.
Long messages are translated sentence by sentence, so a reference across sentences can be lost.
Data
What is saved
The text you type, the English returned, how long it took, and the date and time.
What is not saved
Your name, email, IP address, location, device or browser details. Nothing links a record to you.
Why
Singlish spelling varies more than any generated data can cover. Real typed messages show the researcher what people actually write and where the translator fails.
Who sees it
Only the researcher. Records are stored privately and used for this research.
Removal
Email tishanyasiru@outlook.com with the approximate time and the text, and it is deleted.
Please
Do not type anything private, such as passwords, card numbers or other people's personal details.