-
Notifications
You must be signed in to change notification settings - Fork 14
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.
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
Add first version of time-constrained ORC-WER (tcORC-WER) #52
Conversation
# Add a segment index to the reference so that we can later find words that | ||
# come from the same segment | ||
for i, s in enumerate(reference): | ||
s['segment_index'] = i |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Hmm. Should the assignment be before the filter?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
It shouldn't make a difference because the segment index is here just used for grouping and not for sorting. But it could be easier to understand what's happening when moved before the filter operation
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Hmm. The filter will remove empty segments. So the assignment will not be valid for the input of this function.
Should the assignment also consider the empty segments? Or do we drop the empty segments in the apply assignment function?
We could add this to a ToDo list and solve this later. For the challenge start it is not important.
This PR adds code to compute the tcORC-WER.
Example:
The current version is tested on Libri-CSS with a system that produces 8 streams. It finished computation within 10 minutes and used less than 2GB of RAM (which is a huge improvement over ORC-WER!). These requirements should drop further when the number of streams is smaller.
The code is not fully optimized yet and contains many TODOs. I'll work on some of these TODOs and update the PR during the next few days.
I'll merge main back into this PR once #50 is merged.