non-latin text is not aligned

[beviah]

i.e. arabic, russian, etc.

text field in response has valid non-latin transcript, yet there are no alignments.

[pzelasko]

Can you provide more context?

[beviah]

from lhotse import CutSet, RecordingSet, align_with_torchaudio, annotate_with_whisper
recordings = RecordingSet.from_dir(fld, pattern="*.wav")
cuts = annotate_with_whisper(recordings, device='cuda', language='ar')
cuts_aligned = align_with_torchaudio(cuts)
for cut in cuts_aligned:
    alignments = cut.supervisions

word alignments are always empty for texts of non-latin scripts, as if no text was detected.

[pzelasko]

You’d need an ASR model that supports your target language. You can check if there’s one available in torchaudio: https://pytorch.org/audio/stable/pipelines.html

Otherwise you’d need to train or fine tune your own. You can also try word alignments from faster whisper (unmerged yet in #1017).

[beviah]

Awesome! it works, thanks!