mr-jones123/dka

Automate Low-Resource Language datasets using dka! # dka **Team name:** iForgot **Members:** Xynil Jhed Lacap, Lady Diane Casilang, Raphael Andre Mercado ## AI assistance We used GPT 5.5 through Codex and pi to help build, document, and package this project. ## Links - PyPi: pypi.org - Website: dka.xy-800.workers.dev - Presentation: coast-617.faces.site ## Chosen track We chose **GitHub Education Project Case: Tinig sa Liwanag**. The track asks teams to create reusable, open-source artifacts that advance speech technology for Philippine languages and code-switched speech. Instead of building another app, `dka` focuses on the missing infrastructure layer: preparing raw Philippine-language speech recordings and transcripts into clean, documented, model-ready datasets. ## Problem The Philippines has more than 130 languages, but speech technology support is still uneven. Regional languages such as Cebuano, Hiligaynon, Ilokano, and Waray remain underrepresented in open ASR and TTS tooling. Research shows that Philippine speech datasets are emerging, but the ecosystem still needs reusable preprocessing, quality checks, metadata, and benchmark-ready splits: - The **Philippine Languages Database** paper notes that earlier Filipino speech corpora were often domain-specific, non-parallel, non-multilingual, or insufficient for state-of-the-art ASR/TTS work. - The **UP-DSP Philippine Languages Database** provides 454+ hours across languages including Filipino, Cebuano, Hiligaynon, Ilokano, Waray, and Tausug, but researchers still need tooling to prepare and audit subsets for experiments. - The **iTANONG-DS** paper highlights broader Philippine NLP gaps around benchmark datasets, informal language, geographic variation, and code-switching. ## Solution `dka` is a Python CLI for building Philippine speech datasets. It turns this: ```text raw audio + transcript metadata ``` into this: ```text clean WAV files normalized metadata train/dev/test splits …