Word timestamps not working well for large-v2

#4
by purplerobot - opened

Potentially not designed for large-v2, or maybe bug. Setting wrong (start = end) timestamp for all words in a sequence. Works fine on medium and small

Note that you need to set the correct alignment_heads for the model. More info here: https://gist.github.com/hollance/42e32852f24243b748ae6bc1f985b13a

That did it, huge thanks!

purplerobot changed discussion status to closed
purplerobot changed discussion status to open

Hi again, with the linked timestamps, all works fine on large-v2 with audio >30s, however for all <30s inputs the timestamp is not working

Works fine on same files with medium, but for large-v2 it sets word timestamps to values >= max input length

Sign up or log in to comment