News for hackers. Stuff that matters.
Re:KanjiVG – SVGs of Kanji character strokes including order... by krackers on Tuesday February 21, 2023 @04:47PM (#34887937)
Given a large enough corpus of spoken lines, could you do some ML magic to get the pitch accents (maybe even just FFT and a simple classifier would do)? I'm aware that the "base" pitch accent does change in context so it's not quite trivial, but it seems like you could get pretty close?

Edit: Found https://mizoru.github.io/blog/2021/12/25/Japanese-pitch.html

Re:KanjiVG – SVGs of Kanji character strokes including order... by wahnfrieden on Tuesday February 21, 2023 @05:16PM (#34888408)
The pitch variants are also highly regional but that’s an interesting idea. I wouldn’t want to give it to learners in use cases I can think of, when correct data is available though (just with licensing headaches/costs)
Re:KanjiVG – SVGs of Kanji character strokes including order... by aikinai on Tuesday February 21, 2023 @05:36PM (#34888668)
99% (or more) of the recordings you might find for training would be in the standard accent unless you were specifically digging for regional accents.
Re:KanjiVG – SVGs of Kanji character strokes including order... by wahnfrieden on Thursday February 23, 2023 @12:52PM (#34913928)
it's also changing generationally
Re:KanjiVG – SVGs of Kanji character strokes including order... by gfaure on Tuesday February 21, 2023 @07:45PM (#34889948)
This linked article seems to have gotten special access to the data -- unfortunately only the resulting model is available.
Re:KanjiVG – SVGs of Kanji character strokes including order... by rjh29 on Tuesday February 21, 2023 @09:15PM (#34890690)
It predates modern ML, but that's what Suzuki-kun does:

https://www.gavo.t.u-tokyo.ac.jp/ojad/

Re:KanjiVG – SVGs of Kanji character strokes including order... by krackers on Tuesday February 21, 2023 @09:25PM (#34890771)
No, I don't think Suzuki-kun was trained on voice data, they trained a classifier on an annotated text corpus. And getting access to such a corpus is probably much harder than finding voice samples.

https://research.ibm.com/publications/accent-sandhi-estimati...