Eyyo this is MLo7, mostly post about vocal synth stuff and sometimes singing,,,, still trying tho
Joined March 2023
- Tweets146
- Following7
- Followers81
- Likes35
Aaaand became e2e model
Andromeda is doing so fineee
Training takes long as hell tho and the model is still severely undertrained 😭
Art by wormwood_prod
Been brewing up a vocoder and architecture update. We are using auto-encoder and frequency domain neural vocoder instead.
Music: Mili
ustx: scoliosin
once again mixing sucks but well this is a quick test
Naysan♤ retweeted
「JIZURA 字面」
公開しました
歌詞から文字PV風動画を自動で組み立てるブラウザアプリです。
作例を添付します。こういうものが作れます
是非試してみてください。
852wa.github.io/JIZURA/
Im amazed in how frequency domain neural vocoder solidifies the frequency details way faster than time domain, but both it still take as long to refine into a similar fidelity to one another. Though, I don't even know if different domain has different fidelity ceilings
I wonder what could beat openau whisper features for ASR task to perform auto labeling... its encoder is really good
Been taking a look closer at synthv2 outputs, and I am amazed about how their residuals are so naturally non-coordinated. comparing it to nsf hifigan, they seems to be similar but different. Almost as if sv2 excitations are nonlinearly generated while nsf hifigan is shaped...
SLabeler has been slowly updated
Plan has changed, and I will only make one version instead of separate lite/full version.
I don't want to keep up with two of the same source code because one is already big enough.
m-lo7.itch.io/slabeler/devlo…
