About the dataset structure

#3
by liranzxc - opened

First , great work !

I had question about speakers column at the dataset.
I saw the labels are A,B,1,2.
I looked few examples from the dataset and both have A and B labels in speakers column but A at the first record is not same speaker A at the other record.
We don't need a speaker ID for training? Only care when who speak at single audio file record?

If I want create my own dataset I need to create share labels speakers between records ?

Last question, we training speech diarization on define numbers of speakers ? If yes where it's define?

Thank you very much

TalkBank org
edited May 3

I looked few examples from the dataset and both have A and B labels in speakers column but A at the first record is not same speaker A at the other record.

Each record corresponds to an individual .cha file in the Talkbank database, meaning they are an entirely separate conversation with a different speaker. Best to treat each record seperately.

We don't need a speaker ID for training? Only care when who speak at single audio file record?

The speakerIDs + timestamps are given as separate columns. So, no, we do care about speaker IDs (i.e. tier information).

If I want create my own dataset I need to create share labels speakers between records ?

No. We do not either. Each record (which comes from corresponding .cha file in https://ca.talkbank.org/access/CallHome/eng.html) has a seperate conversation.

Last question, we training speech diarization on define numbers of speakers ? If yes where it's define?

No. If you want, you can check for unique symbols in the speakers column to derive the number of participants for each conversation.

Hope this helps!

I looked few examples from the dataset and both have A and B labels in speakers column but A at the first record is not same speaker A at the other record.

Each record corresponds to an individual .cha file in the Talkbank database, meaning they are an entirely separate conversation with a different speaker. Best to treat each record seperately.

We don't need a speaker ID for training? Only care when who speak at single audio file record?

The speakerIDs + timestamps are given as separate columns. So, no, we do care about speaker IDs (i.e. tier information).

If I want create my own dataset I need to create share labels speakers between records ?

No. We do not either. Each record (which comes from corresponding .cha file in https://ca.talkbank.org/access/CallHome/eng.html) has a seperate conversation.

Last question, we training speech diarization on define numbers of speakers ? If yes where it's define?

No. If you want, you can check for unique symbols in the speakers column to derive the number of participants for each conversation.

Hope this helps!

Thank you ,
If I create my own dataset, that have two differences conversations (two records) , but same person speak on both of them , it's mean the speaker symbol C , must be as labeled in both records as C symbol? I am try to understand if the symbols have meaning for identification the speaker.

Thank you for your time

TalkBank org

No; you do not. Each file is labeled separately. If you want to, however, it wouldn’t hurt as long as the speaker labels are internally consistent.

Sign up or log in to comment