CoolFace
Apppublic

AustinWagner/turkic-transliteration-demo

sourceHugging Faceapache-2.0updated 19d agoView on Hugging Face
1likes
App README

Turkic Transliteration Demo

Overview

This Hugging Face Space demonstrates deterministic broad-IPA transcription for the Turkic language family and a Uralic control language (Finnish). Each language's rule set is derived from a peer-reviewed phonological description of the language, so the output is auditable against the primary literature rather than a black-box model.

Supported languages

  • —Kazakh (kk) — Cyrillic input, IPA output (McCollum & Chen 2021)
  • —Kyrgyz (ky) — Cyrillic input, IPA output (McCollum 2020)
  • —Turkish (tr) — Latin input, IPA output (Zimmer & Orgun 1992)
  • —Azerbaijani (az) — Latin input, IPA output (Ghaffarvand Mokari & Werner 2017)
  • —Uzbek (uz) — Latin or Cyrillic input, IPA output (Ido 2025)
  • —Uyghur (ug) — Arabic-script input, IPA output (McCollum 2021)
  • —Finnish (fi) — Latin input, IPA output (Suomi, Toivanen & Ylitalo 2008)

Kazakh and Kyrgyz additionally support Latin transliteration (Presidential Decree No. 569, 2017; Common Turkic Alphabet, 1991).

Features

  • —IPA Transcription for the full language set above
  • —Latin Transliteration for Kazakh and Kyrgyz
  • —Corpus Download tab: stream text from OSCAR-2301, Wikipedia dumps, and Leipzig-corpus mirrors with optional FastText language-ID filtering
  • —Interactive UI powered by Gradio; no installation required

Usage

  1. 1.Choose a tab: Transliterate to IPA for interactive text conversion, or Download Corpus for batch data ingestion
  2. 2.Select the language from the dropdown
  3. 3.Enter text (or upload a .txt file) and view the transliterated output
  4. 4.Download the result as a UTF-8 text file when it is more than a couple of lines long

Technical Details

This demo is powered by the turkic-translit Python package, which wraps ICU's UTS #35 LDML transform engine (PyICU) with per-language rule files grounded in the phonological literature. The same package also exposes command-line tools (turkic-translit, turkic-download-corpus, turkic-filter-russian) and a Python API for programmatic use.

This Space is a thin wrapper — app.py is ~20 lines and holds no transliteration logic. Everything above comes from the package, pinned exactly in requirements.txt. The pin is exact rather than a floor because Hugging Face reuses a cached wheel when it is given a >= range.

Nothing here is edited by hand. This card, app.py and requirements.txt are all written from `wagner-austin/turkic_transliteration` — the card from .github/hf-space/README.md, the pin from the version in pyproject.toml — so a rule is fixed by pushing there and nowhere else.

Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

Kazakh to IPA Rule set from: Mccollum, Adam & Chen, Si. (2018). Illustration of the IPA: Kazakh. https://www.researchgate.net/publication/328290116IllustrationoftheIPA_Kazakh#fullTextFileContent