VoxCeleb recipe

[desh2608]

I am working on a data preparation recipe for VoxCeleb. The VoxCeleb corpus is used for speaker verification tasks and as such, comes with a standard list of trial pairs. For my own purpose, I am not preparing it for a SV task so I don't need these pairs, but it is reasonable to assume that a number of people would need it. I am not sure how to store the trials list in the manifests, since it is not exactly a "supervision" or a "recording". Any ideas?

[pzelasko]

Yeah that's a good question. I have two ideas:

1. Leave trials as-is and let the users handle it inside speaker ID dataset

2. As part of data prep recipe, for each dev / test split, create two cut-sets where cuts have ids like trial1, trial2, and so on, and contain the original recordings + supervisions. The users can use CutPairsSampler to get trial pairs of cuts for evaluation (note: the dataset has to handle getting a tuple of CutSet as batch).

[desh2608]

The trials file also contains a binary value (0/1) indicating whether the trial target is true or false. What would be a clean way to store these in the associated CutSet pairs --- perhaps using the custom field (#459)?

[pzelasko]

Probably the cleanest way is to save two cut sets — one for positive trials and the other for negative. I don’t thing that using the custom field for this would be clean.

[desh2608]

Cool, thanks!