Skip to content

fix: correct segment order in batched k2 forced_align #76

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
lumpidu wants to merge 1 commit into
base: main
Choose a base branch
from
Open

fix: correct segment order in batched k2 forced_align #76

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
15 changes: 14 additions & 1 deletion src/stylish_tts/train/losses.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -583,6 +583,8 @@ def forced_align(
input_lengths: Tensor,
target_lengths: Tensor,
):
import k2

supervision_segments, token_ids, indices = self.encode_supervisions(
targets, target_lengths, input_lengths
)
Expand All @@ -602,7 +604,11 @@ def forced_align(

best_paths = k2.shortest_path(lattices, use_double_scores=True)
frame_scores = best_paths.scores[(best_paths.labels != -1)]
frame_scores = frame_scores.split(input_lengths.tolist())
# encode_supervisions sorted the batch by length, so best_paths come back
# in that sorted order. Split the scores by the sorted per-segment lengths
# rather than the original input lengths.
sorted_lengths = supervision_segments[:, 2].tolist()
frame_scores = frame_scores.split(sorted_lengths)
scores = torch.stack([p.mean() for p in frame_scores])

batch_arc_shape = best_paths.arcs.shape().remove_axis(1)
Expand All @@ -612,6 +618,13 @@ def forced_align(
# k2 makes an extra frame for some reasons
for i in range(len(batch_frame_labels)):
batch_frame_labels[i][-1] -= 1
# Restore the original batch order so each segment's labels and score line
# up with the caller's segment. Without this every segment receives another
# segment's alignment.
inverse = torch.empty_like(indices)
inverse[indices] = torch.arange(indices.numel(), device=indices.device)
batch_frame_labels = [batch_frame_labels[j] for j in inverse.tolist()]
scores = scores[inverse]
return batch_frame_labels, scores

def on_train_epoch_end(self, train):
Expand Down