speech_recognition/FunASR
1
0
Fork 0
mirror of https://github.com/modelscope/FunASR synced 2025-09-15 14:48:36 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

refine decoding process, merge flow and vocoder

Browse Source
This commit is contained in:
志浩 2024-07-10 16:31:45 +08:00
parent c2c1b66674
commit 04c2d487ed
1 changed files with 6 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
funasr/models/llm_asr/hifigan_module/generator.py
View File

@ -615,7 +615,11 @@ class HiFTGenerator(nn.Module):
real = magnitude * torch.cos(phase) real = magnitude * torch.cos(phase)
img = magnitude * torch.sin(phase) img = magnitude * torch.sin(phase)
inverse_transform = torch.istft( inverse_transform = torch.istft(
torch.cat([real.unsqueeze(-1), img.unsqueeze(-1)], dim=-1), # torch.cat([real.unsqueeze(-1), img.unsqueeze(-1)], dim=-1),
self.istft_params["n_fft"], self.istft_params["hop_len"], self.istft_params["n_fft"], window=self.stft_window) torch.complex(real, img),
self.istft_params["n_fft"], self.istft_params["hop_len"],
self.istft_params["n_fft"], window=self.stft_window,
return_complex=False
)
return inverse_transform.unsqueeze(-2) # unsqueeze to stay consistent with conv_transpose1d implementation return inverse_transform.unsqueeze(-2) # unsqueeze to stay consistent with conv_transpose1d implementation