Help on the "proper way" to build a protocol for DIHARD database with conflicting URI

[pyannote]

Help on the "proper way" to build a protocol for DIHARD database with conflicting URI

[hbredin]

I noticed the same issue when working on DIHARD challenge. I would argue that this is bad practice from DIHARD organizers to use the same name for different files... but now the harm is done :)

The solution based on the additional split field should work.
I don't foresee any issue with the rest of pyannote ecosystem (though I might be wrong because I haven't had my morning coffee yet...)

An alternative is to use /path/to/dihard_data/{uri}.flac as matching pattern and use dev/data/single_channel/flac/DHxxxx as actual uri. This might cause an issue with pyannote-metrics.py evaluation script. In particular, get_hypothesis should probably be updated to support this kind of uri mismatch: https://github.com/pyannote/pyannote-metrics/blob/efa42b99f420210e4eb0225a1be972718b0a6bc2/scripts/pyannote-metrics.py#L142-L181

[hadware]

Well, i'm going to try the 'split' method out and see how it goes.

Regarding the alternative method, I gazed intensely at the code but couldn't see anything that would be disturbed by having subpaths as uri's.

[geniki]

In our case at least, I haven't seen the benefit of having the automatic FileFinder over using fixed paths within a dataset. As a downside, it seems to take some time to find all files for a bigger dataset, slowing startup time.

[hadware]

Just so you know, we decided (someone else from CoML) to just remodel the dataset with a script and use a good ol' "Generic speaker diarization protocols" to save us the hassle of having to implement an actual DB module.

[hbredin]

FYI, I am currently working on v2 of pyannote.audio where you will be able to use your own piece of code in place of FileFinder, by overriding the "audio" preprocessors.

I am also playing with lazy loading in pyannote.database so that training starts as soon as possible (without looking for all files first). If you want to try it (and provide me with feedback), simply checkout and install lazy_preprocessors branch of pyannote.database.

[hbredin]

Closing in preparation of pyannote.database 4.0 which comes with a complete rewrite of custom protocols. See custom branch for details.