Description
@hhadian, this is another one for you, which I'd like going to the svd_draft branch in my personal repo.
It's something I want for the rewrite of chain scripts to make it easier to use mismatched egs that were previously dumped.
The job is to extend nnet3-chain-copy-egs, to supply at least a minimum context in input features, by duplicating frames as needed. E.g. if you supply the option --extend-left-context=12 --extend-right-context=10, and it notices that the left-context is less than 12 and/or the right-context is less than 10, it will extend the eg with repeats of the first and/or last frame. Note: I think we can assume that the order of sequences in the egs is: all frames of sequence n=0; all frames for n=1; and so on, since that's how nnet3-chain-merge-egs merges them, I believe. Also please change the current --left-context and --right-context options so that instead of just truncating the input context, they will extend it by duplicating frames. This won't break any existing scripts. And have it print some kind of summary at the end if something was done so that we know what percentage of egs were extended for, say, left-context, and for those that were extended, by how many frames on average. (And the same for right-context). If it turns out to be easier to also make the same changes for nnet3-copy-egs, doing it that way is fine.