What is Script Split?
One spoken language written in two different scripts, with Hindi in Devanagari and Urdu in Perso-Arabic as the standard case.
A script split means the same audio yields two mutually incompatible text corpora. Which script the project uses has to be settled before work starts.
Serbo-Croatian, Kazakh (Cyrillic to Latin) and Mongolian (Cyrillic and traditional Mongolian script) are the other standard cases.
Where script splits occur
The classic pattern is one spoken language written in two systems. Hindi and Urdu share their spoken core and are written in Devanagari and Perso-Arabic; a transcript in one is not readable as the other. Punjabi splits the same way between Gurmukhi and Shahmukhi, and Serbian is written in both Cyrillic and Latin, with the same words available in either.
A second form is romanization: writing a language in Latin letters informally. Romanized Hindi and the Arabic chat alphabet are everyday examples, produced by people typing quickly on phones. Transcripts collected from real users often arrive in this form even when the project's formal script is something else, so the brief has to say which one is wanted.
The third is script reform, where a country changes its official writing system. Kazakh is moving from Cyrillic toward Latin, and Mongolian uses Cyrillic and its traditional script side by side. Material collected before a reform and after it can be in different scripts for the same language, which matters when old and new data are pooled.
What to specify when buying
Fix the script in the brief before annotation starts. If the brief names the language but not the script, annotators from different backgrounds will deliver a mix, and a mixed batch is worse than either pure version: the same word written in two scripts becomes two different tokens, and the model splits its learning between them instead of reinforcing it.
Ask whether transliteration is included, and in which direction. Converting a transcript between scripts is a defined task with tools, but it is not free and it is not perfect; names and loanwords are where conversion breaks. If the deployment will see romanized text, training on the formal script alone leaves a gap, and the brief should say how that gap is covered.
Check what the audio actually supports. Where the two scripts belong to different registers, as Hindi and Urdu largely do in formal vocabulary, the same recording will not produce the same words in both. The script decision is therefore also a vocabulary decision, and it should be made by the buyer, not left to the annotation team.
Related terms
-
Transcription
Writing down what is said in a recording. It is the core annotation task in any speech dataset.
-
Annotation
Attaching machine-readable labels to raw data, a transcript, an intent, a speaker identity.
Keep reading
-
Full glossary
Every term we explain, in one list.
-
How to buy training data
Where these terms actually show up, and which ones change a quote.
-
Help center
How a project runs, from specification to delivery.
Submit a sourcing request
Tell us the language, the hours, and what the data needs to look like. You will get a real number and a real timeline, not a range. If we cannot source it well, we will tell you that instead.
- Pilot batch before the full run, so problems surface early.
- Consent documentation delivered with the data.
- No medical or clinical data. No recorded telephone calls.
- You send the spec Language, hours, recording conditions, and the deadline.
- We reply within two business days A real number and a real timeline, not a range.
- You approve a pilot batch A small run first, so problems surface before the full order.
- Full run, then delivery with documents Audio, transcripts, and the consent record for every speaker.