kalixlouiis/en-my-written-spoken-parallel

English-Burmese Written and Spoken Parallel Dataset

An advanced, human-curated English-to-Burmese parallel dataset specifically designed to explicitly split Burmese translations into two distinct linguistic registers: Written Style (ရေးဟန် / Literary) and Spoken Style (ပြောဟန် / Colloquial). The content is predominantly focused on the technology and global news domains.

• —Creator: Khant Sint Heinn
• —Project Status: ⚠️ Active & Ongoing (Work in Progress) — This dataset is actively curated, and new data will be added continuously over time.

Dataset Description

Introduction to Burmese Registers (Written vs. Spoken)

In the Burmese language, the distinction between "Spoken Style" (ပြောဟန်) and "Written Style" (ရေးဟန်) is deeply rooted, critical, and often overlooked in standard NLP applications.

• —Spoken Style (ပြောဟန်): A natural, colloquial register used in spoken conversations, podcasts, personal letters, dialogues in literature, speeches, and conversational articles/blogs. It fosters intimacy and is significantly easier to understand in daily interactions.
• —Written Style (ရေးဟန်): A formal, disciplined, and literary register mandatory for official news broadcasts, academic textbooks, research papers, legal documents, and corporate records.

The Linguistic Difference

The structural divergence between the two forms does not lie within the core vocabulary, but heavily within the grammatical particles (post-positional markers and verb suffixes).

• —Affirmative Sentence Example:
• —English: "Maung Tint came down from Mandalay to Yangon by train."
• —Spoken (ပြောဟန်): မောင်တင့်ဟာ မန္တလေးကနေ ရန်ကုန်ကို ရထားနဲ့ ဆင်းလာတယ်။
• —Written (ရေးဟန်): မောင်တင့်သည် မန္တလေးမှ ရန်ကုန်သို့ ရထားဖြင့် ဆင်းလာသည်။ (Notice how core nouns/verbs like Mandalay, Yangon, and train remain unchanged, while particles swap entirely: ဟာ ➔ သည်, ကနေ ➔ မှ, ကို ➔ သို့, နဲ့ ➔ ဖြင့်, and တယ် ➔ သည်).

• —Interrogative Sentence Example:
• —English: "Why didn't he come to school?"
• —Spoken (ပြောဟန်): သူ ဘာလို့ ကျောင်းမလာတာလဲ။
• —Written (ရေးဟန်): သူ အဘယ်ကြောင့် ကျောင်းမလာသနည်း။

• —Negative Sentence Example:
• —English: "He did not come to school today."
• —Spoken (ပြောဟန်): သူ ဒီနေ့ ကျောင်းမလာဘူး။
• —Written (ရေးဟန်): သူ ယနေ့ ကျောင်းမလာပါ။

Why This Dataset Matters

Most commercial machine translation engines flatten Burmese into a single form or hallucinate hybrid registers, resulting in text that feels either unnaturally robotic or inappropriately casual. By training models on explicitly paired written and spoken paths, AI agents can successfully perform context-aware style transfers, natural dialogue generation, and flawless localization.

Dataset Structure

The dataset is structured as a single clean parallel corpus with three major features:

1. 1.`English`: The original source sentence (primarily tech news and media text).
2. 2.`Burmese_written`: The formal news/literary translation (formal, active voice, structurally accurate).
3. 3.`Burmese_spoken`: The natural conversational translation (active voice, easy to understand, human-like rhythm).

Data Instance Example

{
  "English": "The latest instalment in Rockstar's blockbuster game franchise, Grand Theft Auto, is set to be the biggest games launch of the year.",
  "Burmese_written": "Rockstar ကုမ္ပဏီ၏ အောင်မြင်ကျော်ကြားလှသော Grand Theft Auto (GTA) ဂိမ်းစီးရီးမှ အသစ်ထွက်ရှိလာမည့် ဂိမ်းသည် ယခုနှစ်အတွင်း အကြီးမားဆုံး ဂိမ်းမိတ်ဆက်ပွဲ ဖြစ်လာရန် ရှိနေပါသည်။",
  "Burmese_spoken": "Rockstar ကုမ္ပဏီရဲ့ အောင်မြင်ကျော်ကြားလှတဲ့ Grand Theft Auto (GTA) ဂိမ်းစီးရီးရဲ့ အသစ်ထွက်ရှိလာမယ့် ဂိမ်းဟာ ယခုနှစ်အတွင်း အကြီးမားဆုံး ဂိမ်းမိတ်ဆက်ပွဲ ဖြစ်လာဖို့ ရှိနေပါတယ်။"
}

Data Curation and Domain Focus

• —Primary Domain: Technology News, Software/Gaming Announcements, Artificial Intelligence, and Modern Digital Culture.
• —Translation Quality: Purely human-curated and polished. It filters out raw machine-translation artifacts, enforcing active-voice patterns and culturally resonant phrasing that reads seamlessly to a native Burmese speaker.

Project Status & Updates

This repository is not static. It serves as an active repository for a running, processing project. The data pipeline is ongoing, and the dataset size will be progressively scaled up with regular commits. If you are using this for benchmarking or model training, make sure to pull the latest versions.

Licensing & Contributions

This dataset is licensed under the Creative Commons Attribution 4.0 International (CC-BY-4.0) license. You are free to share and adapt the material for any purpose, even commercially, as long as appropriate credit is given to the creator.

Contributions, corrections, or expansions targeting specific domains are highly welcome. Feel free to open an issue or reach out to the project creator.

• —Created and Maintained by: Khant Sint Heinn (Kalix Louis)