Why 0.667 during inference in vits?

[p0p4k]

TTS/TTS/tts/models/vits.py

Line 569 in 1e15269

inference_noise_scale: float = 0.667

TTS/TTS/tts/models/vits.py

Line 1155 in 1e15269

z_p = m_p + torch.randn_like(m_p) * torch.exp(logs_p) * self.inference_noise_scale

After training, there is possibility that the distribution we want to sample from does not completely match the z_spec. Also, to maximize the possibility of high probable values we scale the variance and sample from a smaller Gaussian, thus leading to better empirical sampling.