update zero mean gaussian noise

README.md

@@ -8,9 +8,14 @@ Implementation of the [CoMospeech](https://arxiv.org/pdf/2305.06908.pdf). For al
 
 # Update
 
+
+
+We find that zero-mean Gaussian noise instead of the prior in grad-tts can also achieve similar effects. We alse release the new code and checkpoints.
+
 We add   Heun’s 2nd order method support for   teacher model. (can be used for teacher model sampling and better ODE trajectory for consistency distillation)
 
 
+
 ## Abstract
 
 **Demo page**: [link](https://comospeech.github.io/).
@@ -51,4 +56,5 @@ Checkpoints trained on LJSpeech can be download from [here](https://drive.google
 I would like to extend a special thanks to authors of Grad-TTS, since our code base is mainly borrowed from  [Grad-TTS](https://github.com/huawei-noah/Speech-Backbones/blob/main/Grad-TTS/).
 
 ## Contact
-You are welcome to send pull requests or share some ideas with me. Contact information: Zhen YE ( zhenye312@gmail.com )
+You are welcome to send pull requests or share some ideas with me. Contact information: Zhen YE ( zhenye312@gmail.com )
+

model/como.py

@@ -247,8 +247,8 @@ def EDMPrecond(self, x, sigma ,cond,denoise_fn,mask):
 
         sigma = sigma.reshape(-1, 1, 1 )
 
-        c_skip = self.sigma_data ** 2 / (sigma ** 2 + self.sigma_data ** 2)
-        c_out = sigma * self.sigma_data / (sigma ** 2 + self.sigma_data ** 2).sqrt()
+        c_skip = self.sigma_data ** 2 / ((sigma-self.sigma_min) ** 2 + self.sigma_data ** 2)
+        c_out = (sigma-self.sigma_min) * self.sigma_data / (sigma ** 2 + self.sigma_data ** 2).sqrt()
         c_in = 1 / (self.sigma_data ** 2 + sigma ** 2).sqrt()
         c_noise = sigma.log() / 4