Comments (5)
An initial attempt:
- Conv1dResnet + skip connection https://soundcloud.com/r9y9/20220309-nit-song070-svs-world-conv-sine-based-vibrato-modeling-skip-connection-nitech-jp-song070-f001-003
I think the pitch trajectory gets smoother compared to the following samples:
- Conv1dResnet: https://soundcloud.com/r9y9/20220308-nit-song070-svs-world-conv-vibrato-modeling-nitech-jp-song070-f001-003
- Conv1dResnetMDN (w/ vibrato modeling): https://soundcloud.com/r9y9/20201116-nit-song070-svs-world-conv-mdn-dim-wise-mdn-nitech-jp-song070-f001-003
from nnsvs.
Here's a prototype of Conv1dRenet + skip connection for the record:
class ResConv1dResnet(BaseModel):
def __init__(
self,
in_dim, hidden_dim, out_dim, num_layers=4,
in_lf0_idx=300,
in_lf0_min=5.3936276,
in_lf0_max=6.491111,
out_lf0_idx=180,
out_lf0_mean=5.953093881972361,
out_lf0_scale=0.23435173188961034,
):
super().__init__()
self.in_lf0_idx = in_lf0_idx
self.in_lf0_min = in_lf0_min
self.in_lf0_max = in_lf0_max
self.out_lf0_idx = out_lf0_idx
self.out_lf0_mean = out_lf0_mean
self.out_lf0_scale = out_lf0_scale
model = [
nn.ReflectionPad1d(3),
WNConv1d(in_dim, hidden_dim, kernel_size=7, padding=0),
]
for n in range(num_layers):
model.append(ResnetBlock(hidden_dim, dilation=2 ** n))
model += [
nn.LeakyReLU(0.2),
nn.ReflectionPad1d(3),
WNConv1d(hidden_dim, out_dim, kernel_size=7, padding=0),
]
self.model = nn.Sequential(*model)
def forward(self, x, lengths=None):
out = self.model(x.transpose(1, 2)).transpose(1, 2)
# denormalized lf0 from the input musical score
lf0_score = x[:, :, self.in_lf0_idx].unsqueeze(-1)
lf0_score_denorm = lf0_score * (self.in_lf0_max - self.in_lf0_min) + self.in_lf0_min
# TODO: must be careful about dynamic features
# Residual connection in denormalized f0
lf0_res = out[:, :, self.out_lf0_idx].unsqueeze(-1)
lf0_res = 0.693 * torch.tanh(lf0_res)
lf0_pred_denorm = lf0_res + lf0_score_denorm
# Back to normalized f0
lf0_pred = (lf0_pred_denorm - self.out_lf0_mean) / self.out_lf0_scale
out[:, :, self.out_lf0_idx] = lf0_pred.squeeze(-1)
return out
from nnsvs.
PR is up #79. A heuristic parameter 0.693
was replaced with a better value. Also added some more comments in the code. Will add the sinsy's acoustic model soon.
Here's the distribution of residual log-F0 for nit-song070 database:
The most of data (>99.7%) is in the range of [-0.35 ~ 0.35] (i.e. [-600, 600] (in cent))
from nnsvs.
Now I merged #73. Next, I will revise my local implementation for the new acoustic model and make a PR soon.
from nnsvs.
fixed by #81
from nnsvs.
Related Issues (20)
- Improved duration modeling with relative note duration prediction HOT 1
- 想制作中文数据库 HOT 1
- assert exists(wav_path) AssertionError HOT 1
- Configs for 44.1 kHz
- Incorporate the uSFGAN training step into recipes
- Improvements related to NNSVS paper
- Add DiffSinger configurations and recipes to reproduce experiments reported in NNSVS documentation HOT 6
- Diffusion-based acoustic models HOT 2
- The combination of NPSSMDNMultistreamParametricModel and BiLSTMResF0NonAttentiveDecoder with use_mdn=True results in training failure HOT 4
- Converting Enunu to NNSVS HOT 3
- A Question HOT 1
- Refactor svs.py to be more modular and extensible for ENUNU
- Remove the trainable post-filter functionality to make code simple
- A separate training script for F0 prediction model HOT 1
- Can nnsvs be run on AMD GPUs via ROCm? HOT 2
- Using Enunu english model on NNSVS HOT 3
- AttributeError: module 'matplotlib' has no attribute 'axes'
- Is it unnecessary to resample the audio which is not 48k to 48k? HOT 2
- !pip install nnsvs HOT 6
- Cannot install nnsvs HOT 2
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from nnsvs.