zceng / lvcnet Goto Github PK

hello @zceng, thanks for your contribution in LVCNet.
I follow your default experiment's setting about training LJSpeech. Including sampling rate = 22050, batch_size = 8 and so on. But I have a poor synthesis although training reached checkpoint-180000.

real:

generate:

wavs zip:
0100_wavs.zip

Is it because not enough training? or other reasons?
Looking forward to your reply.
Thanks.

LVCnet他没有nsf哈斯会有hifigan一样的问题 （容易断音）
但是lvc 不是直接上采样mel 还有方法可以添加nsf吗？

nsf 就是下面链接的那种
https://github.com/vtuber-plan/NSF-HiFiGAN
https://nii-yamagishilab.github.io/samples-nsf/

How may days needed for training？ And what is the batch_size？

In MELGAN, there is not input noise as x. So

首先，感谢您在声码器上做的努力
我想问一下，推理速度上有与原始LPCNet做过对比吗

https://github.com/ZENGZHEN-TTS/LVCNet/blob/6e7748e7ef358f7d95b8ce08bb682e7993f9c639/vocoder/models/lvcnet.py#L76

https://github.com/ZENGZHEN-TTS/LVCNet/blob/6e7748e7ef358f7d95b8ce08bb682e7993f9c639/vocoder/models/lvcnet.py#L45

if I pad 2

wavs: 16k hop_size=200
when i set:
generator_params:
in_channels: 1 # Number of input channels.
out_channels: 1 # Number of output channels.
inner_channels: 8
cond_channels: 80
cond_hop_length: 200 #256
lvc_block_nums: 3
lvc_layers_each_block: 10
lvc_kernel_size: 3
kpnet_hidden_channels: 64
kpnet_conv_size: 1
dropout: 0.0
use_weight_norm: true # Whether to use weight norm.

When my data is 16k, hop_size=200, how should the parameters be set? thank you