mrq/vall-e
5
9
Fork 0
Code Issues 8 Pull Requests Packages Projects Releases Wiki Activity

moved duration padding for NAR-len to be a scalar instead (since it seems longer utterances need it much more so than shorter utterances)

Browse Source
This commit is contained in:
mrq 2024-11-21 13:04:07 -06:00
parent 6aee08f9c0
commit 2a084544e8
1 changed files with 2 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
vall_e/inference.py
View File

@ -272,13 +272,13 @@ class TTS():
with torch.autocast("cuda", dtype=self.dtype, enabled=self.amp):
if model_len is not None:
# extra kwargs
duration_padding = sampling_kwargs.pop("duration_padding", 1)
duration_padding = sampling_kwargs.pop("duration_padding", 1.05)
nar_len_prefix_length = sampling_kwargs.pop("nar_len_prefix_length", 0)
len_list = model_len( text_list=[phns], proms_list=[prom], task_list=["len"], disable_tqdm=not tqdm, **{"max_duration": 5} ) # don't need more than that
# add an additional X seconds
len_list = [ l + duration_padding * cfg.dataset.frames_per_second for l in len_list ]
len_list = [ l * duration_padding * cfg.dataset.frames_per_second for l in len_list ]
kwargs = {}
# nasty hardcode to load a reference file and have that as the input target