完整歌曲效果

[MaxMax2016]

reg_pit_song.mp4

F0根据规则生成，没有生成颤音曲线，机器人感觉严重

[dutchsing009]

Hello , good thing you opened the repository again i emailed you about it 2 days ago. The audio above sounds so robotic and not natural at all. can you tell me exactly what is the problem here so i see if i can solve it ??
regarding the other issue #3 when you said "but need more detail of f0" i think i have a solution for you which is...
you need to extract the Fundamental frequency from the audio during preprocessing , you extract that from the waveform using this https://github.com/NVIDIA/mellotron/blob/master/yin.py
compute_yin() computes Fundamental frequency from the waveform
w_step should be the size as STFT hop size

those parameters may need to be changed for SVS since f0 can be higher than in TTS
that was my recommendation for you regarding the #3, please tell me if that works ? or if there are any other problems you also have especially regarding the so robotic audio above :)
Best Regards

[MaxMax2016]

I mean need dnn network to predict smooth f0 with vibrato; The audio above sounds so robotic ,because this f0 is producted by handed(midi map) withnot vibrato; or use https://github.com/stakira/OpenUtau draw smooth f0 with vibrato;

[dutchsing009]

ah ok so now you are back to issue #1 again that you need vibrato for it not to have robotic sound.

[MaxMax2016]

@dutchsing009 i use nn mode to predict the F0, the song is here #8