Taigi speech recognition
This page is about the speech half: why the frontline needs a model that understands Taigi, which step MediaTek had already done, which step we filled in, and the one we tried and haven't got right yet. Three of the four files we have on Hugging Face come from this work.
Why it's needed
A care worker sits beside an older adult at a community station, tablet in hand. Did you sleep well last night? Do you get up in the middle of the night? The older adult answers. The worker has to listen, translate the answer into whatever the form wants, and type — all at once, head down. The older adult ends up talking to the top of their head.
And the older adult is speaking Taigi. Often Mandarin and Taigi land in the same sentence: "my leg, it goes sng (aches)." The worker understands it perfectly. They just have to convert it to Mandarin before they can type it.
What we wanted was simple: let them look up, talk to the person properly, and have the record write itself. That takes a speech model that understands Taigi — and off-the-shelf speech recognition handles Mandarin well and Taigi badly. Older adults are not going to switch to Mandarin to suit our software.
Someone in Taiwan had already built a model that understands Taigi. What we filled in was "it doesn't fit on the computer that's actually in the room."
Which step is ours
Breeze-ASR-26 was trained and open-sourced by MediaTek Research (Apache-2.0, derived from OpenAI Whisper large-v2), and it understands Taigi and Taiwanese accents. We did not train it — that needs saying plainly.
But it is 3.1 GB and wants a GPU to run smoothly. The six-year-old PC at a community station cannot hold it, and a care worker's phone certainly cannot. So we did the part that was left: shrink it, actually measure the speed and the memory, build three versions for three kinds of machine, and put them back online.
Three builds
The only difference is what hardware they run on. All Apache-2.0 — commercial use and modification without asking us.
The fast one, if you have a server
Works on one old laptop — just drag the file in
Taigi recognition on a phone or in a browser
On speech longer than five seconds you cannot tell the compressed version from the original (about 3.6% character error against it). On a home visit a phone is often the only device there is, which is why the slowest build had to exist too.
The one we didn't publish
We also tried shrinking it to 237 MB so it could run on a phone directly. There was real progress: on genuine long-form speech, the gap to the original model closed from 63% to 38.7%. But the original model itself is at 14%. Still too far, so we did not publish it.
One failure along the way is worth writing down. On the first attempt we pushed the small model too hard to imitate the big one's manner, and it got worse. Progress only came after we put the right answers back in charge and let imitation play a supporting role. We publish records like that too, because they help other people more than a nice-looking number does.
What we are stuck on is not training technique. It is that there is very little recorded Taigi. Every open source added together comes to a few dozen hours, mostly short sentences read off a script — a long way from how an older adult actually talks.
There is a more fundamental problem too. The faithful way to write spoken Taigi is Tâi-lô romanization; writing it in Mandarin Han characters throws the Taigi reading away. Written in Han characters, 斷腦筋 gives the reader no clue that it means a stroke. So what we are building now outputs Tâi-lô. It isn't ready. When it is, it will show up on this page.
What it does in the product
What comes out of speech is still how the frontline talks. 皮蛇 is still 皮蛇; it has to map back to 帶狀皰疹 before it can be stored — and that is a different model's job. Only with both halves does a spoken sentence become a usable record.
Someone in Taiwan had already built a model that understands Taigi. We made it fit on the computer that's actually in the room.
All three builds are on Hugging Face under Apache-2.0. Each model card carries the measurement method, the limits, and what it is suited for. The upstream model is MediaTek Research's Breeze-ASR-26 — they trained it; we quantized, measured, and converted it.