11 ms·
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 awa
by krackers 4y ago
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 https://mizoru.github.io/blog/2021/12/25/Japanese-pitch.html
- wahnfrieden 4y agoThe 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) The correct data already exists, so I'm not sure what the point is besides having a less accurate but freer option
- aikinai 4y ago99% (or more) of the recordings you might find for training would be in the standard accent unless you were specifically digging for regional accents.
- wahnfrieden 4y agoit's also changing generationally
- gfaure 4y agoThis linked article seems to have gotten special access to the data -- unfortunately only the resulting model is available.
- rjh29 4y agoIt predates modern ML, but that's what Suzuki-kun does: https://www.gavo.t.u-tokyo.ac.jp/ojad/ https://www.gavo.t.u-tokyo.ac.jp/ojad/
- krackers 4y agoI 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-estimation-of-tokyo-dialect-of-japanese-using-conditional-random-fields https://research.ibm.com/publications/accent-sandhi-estimati...