Incorrect speaker id fields

#5
by akashmjn - opened

Hi team,

I've noticed incorrect values for speaker ids in this dataset.

See below an example from split (microphone-single, validation) dataset[0].
Both word_speakers and segment_speakers only contain the value '6'.

image.png

The channels field however, still appears to correspond to speaker IDs. So while not a blocking issue, it'll likely cause some confusion.

(cc @albertvillanova noticed your recent AMI-related activity )

hi @akashmjn !
this version of the AMI dataset should actually be deprecated, you can find a correct one here: https://huggingface.co/datasets/edinburghcstr/ami

I'm gonna add the deprecation warnings to this repo, thank you for reminding me :D

polinaeterna changed discussion status to closed

Hi @polinaeterna ,

Thanks for your response!

Actually, I did take a look at edinburghcstr/ami - and it appears that it's only available in pre-segmented (chunked) form?

AFAIK ami was the only dataset that contained meetings as full sessions. Flexibility to create segments from multiple speakers is still needed e.g. for diarization one needs to create multi-speaker segments. See this project for a reference on how the data is prepared.

What do you recommend? (for now, i'll continue to use ami as I can still see speaker info preserved after running load_dataset)

Best,
Akash

Sign up or log in to comment