Proper nouns
People, places, schools, government agencies, news terms, and local systems are often replaced with versions from outside Taiwan or pronounced incorrectly.
Open foundation models for Taiwan's language contexts
Let AI speak Taiwan,hear Taiwan,and read Taiwan.
General AI models often miss usage in Taiwan, local accents, Taigi terms, proper nouns, and sentences that mix Chinese, English, and Taigi. OpenFormosa builds open models and tools for these real scenarios, with lower deployment cost and room for fine tuning.
Try it
Hear how the demo handles mixed Chinese and English sentences, Taiwan proper nouns, and natural speech.
Try the TTS demo
Run it
Scroll down for the released tokenizer, base model, TTS demo, and upcoming work.
See model and technical route
Join in
Share test sentences, documentation issues, data licensing leads, or integration examples.
Join Discord
Models and technical route.
Remember Taiwan
Base · Barbet
A shared base for Taiwan context, long documents, culture, and knowledge.
Speak Taiwan
TTS · Blue Magpie
Natural Taiwanese Mandarin voice with Bopomofo and Tâi-lô support.
Hear Taiwan
ASR · Black Bear
Taiwan Mandarin, Taigi and Hakka accents, and speech that mixes Chinese and English.
Read Taiwan
OCR · Sika Deer
Layout and structure of public documents, receipts, forms, menus, and addresses.
- Shippedweights, a demo, or a report you can open now
- In progressactively training or building
- Plannedon the roadmap, not started yet
-
PangolinTokenizer Tokenizer Shipped
A byte level BPE tokenizer built for Taiwan. It has 114,688 merges and reduces tokens per character on PangolinBench with a smaller vocabulary, so the same context window can hold more real Taiwan content.
-
Barbet 1B Base Base Shipped
A 1B hybrid language model with long context support. It serves as the shared Taiwan context base for ASR, TTS, and OCR tasks. The weights, Hugging Face conversion, and full technical report are available.
-
BlueMagpie-TTS TTS Shipped
Speech synthesis for Taiwanese accents and mixed Chinese and English text. On a hard pronunciation set, it reduces character error rate by about 58% against the baseline, which means less manual correction for news voiceover, exhibition guides, and educational content. The online demo is live.
-
BlackBear-ASR ASR Planned
A planned ASR branch for Taiwan Mandarin, Taigi and Hakka accents, speech that mixes Chinese and English, meetings, and classroom audio. It has not been trained yet and has no weights or demo.
-
SikaDeer-OCR OCR Planned
A planned OCR branch for Taiwan receipts, forms, tables, menus, addresses, and archive documents. It has not been trained yet and has no weights or demo.
Why this matters
Writing Traditional Chinese does not mean understanding usage, accents, and culture in Taiwan.
News voiceover, exhibition guides, classroom transcripts, Taiwan addresses, institution names, Taiwan Mandarin, Taigi and Hakka accents, Bopomofo, Tâi-lô, Taigi words, and sentences that mix Chinese and English often need manual correction. OpenFormosa helps smaller models handle tasks that used to rely on large models or manual work, lowering deployment cost and making more Taiwan content easier to generate, transcribe, read aloud, and rewrite.
Taiwan scenarios where generic models often fail.
Chinese, English, and Taigi mixed text
Everyday language in Taiwan often puts Chinese, English abbreviations, Taigi words, and domain terms in the same sentence. Generic models often split those boundaries incorrectly.
Accents, tone, and narrated content
Taiwan Mandarin, Taigi and Hakka accents, speech that mixes Chinese and English, news scripts, exhibition guides, and educational content need local pronunciation, pauses, and tone.
Documents and layout
Official letters, invoices, receipts, forms, menus, temple lots, and Taiwan address formats need OCR and structural understanding that generic models often miss.
Who OpenFormosa is for.
Education and research institutions
- Classroom transcripts and lecture summaries
- Educational TTS and learning materials
- Language, culture, and model evaluation research
Developer contributors
- Improve docs, Colab, and Quickstart
- Report test sentences, errors, and issues
- Help with model loading, examples, and PRs
Partner organizations
- Media: podcasts, video, and subtitles
- Venues: exhibition guides and OCR for archives
- Public sector: document summaries and secure deployment
Software integrators
- Customer service ASR and TTS
- Meeting notes, document OCR, and form workflows
- Internal knowledge base RAG and private deployment
Open source is a trust mechanism
We open more than model weights. We also open evaluation methods.
Open source is how we build trust. Beyond model weights, we publish evaluations designed for Taiwan, benchmark systems, model cards, training lessons, and core implementation work.
Read evaluation workflow- Try directly BlueMagpie-TTS has an online demo for usage in Taiwan, mixed Chinese and English sentences, and proper nouns.
- Run locally Barbet, PangolinTokenizer, and related technical reports provide weights, settings, and reproducibility notes.
- Check the numbers Benchmarks should include test sets, baselines, tasks, and real usage differences, not only scores.
- Help close gaps Docs, Quickstart, Colab, error cases, test sentences, and data licensing leads can all become contributions.
OpenFormosa is an open family of AI models. With the Gaji bag (Taiwan Striped Bag) as its symbol, it weaves Taiwan's voices, texts, images, and memories into model capabilities people can try, inspect, and deploy.
If you work with Taiwan language, speech, documents, or local use cases, join us on GitHub or in the community to discuss testing, docs, model integration, data licensing, and collaboration. If you only want to see the results, start with the demo and released models.
Make local usage, local accents, and mixed language contexts part of everyday model work.
Read the project idea