SapinSapin AI — Open Foundations for Philippine AI

SapinSapin AI builds open speech and language foundations so Philippine AI can be made with — and for — the people who speak it.

What We Build

Open datasets, speech recognition models, text-to-speech systems, and language models for 10+ Philippine languages: Filipino, Cebuano, Ilocano, Hiligaynon, Bikol, Pangasinan, Kapampangan, Tausug, Waray, and English.

Datasets

Philippine Language Dataset (PLD) — 334,268 utterances across 10 languages, 448+ hours of prompted 16 kHz speech
Filipino Speech Corpus — 305,246 segments, 65+ hours of Filipino/Tagalog speech (MIT license)
BantayWika — 6.09M verified tokens of Filipino, Cebuano, and Ilocano text
halohalo — 45,536 rows of cleaned Philippine-language web data

Models

28 public models spanning speech recognition (Whisper fine-tunes), text-to-speech (SpeechT5 fine-tunes), voice conversion, and text generation (Llama 3.1 and GPT-oss-20b continued pretraining).

Live Demo

Try speech recognition, text-to-speech, and voice conversion for Philippine languages at the halohalo Space on Hugging Face.

FAQ

Which license applies?: Licenses are dataset-specific. MIT and the UP-DSP research license appear among the current datasets. Always read the dataset card before use.
Can I use the data commercially?: It depends on the individual dataset license and any access conditions. The linked Hugging Face dataset card is the source of record.
How do I contribute?: Open an issue or discussion on a SapinSapin repository on GitHub.
What is the model roadmap?: The catalog includes language, speech recognition, text-to-speech, and audio-to-audio models. Follow the Hugging Face activity for updates.