mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-12-11 06:55:27 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fixed the issue of errors in fully silent sentences during evaluation.

Browse Source
This commit is contained in:
drawfish 2023-05-22 16:46:15 +08:00 committed by GitHub
parent 7c4ff66a3d
commit e49fe2e7c4
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 6 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/frame_reducer.py
View File

@ -74,9 +74,13 @@ class FrameReducer(nn.Module):
padding_mask = make_pad_mask(x_lens) padding_mask = make_pad_mask(x_lens)
non_blank_mask = (ctc_output[:, :, blank_id] < math.log(0.9)) * (~padding_mask) non_blank_mask = (ctc_output[:, :, blank_id] < math.log(0.9)) * (~padding_mask)
if y_lens is not None: if y_lens is not None or self.training == False:
# Limit the maximum number of reduced frames # Limit the maximum number of reduced frames
if y_lens is not None:
limit_lens = T - y_lens limit_lens = T - y_lens
else:
# In eval mode, ensure audio that is completely silent does not make any errors
limit_lens = torch.ones_like(x_lens)
max_limit_len = limit_lens.max().int() max_limit_len = limit_lens.max().int()
fake_limit_indexes = torch.topk( fake_limit_indexes = torch.topk(
ctc_output[:, :, blank_id], max_limit_len ctc_output[:, :, blank_id], max_limit_len