Several questions about open_stt

[snakers4]

Hi,

(1)
Did not track
We do not have proper labels for speakers in this dataset

(2)
There is an issue where I posted a bench on the dataset with CER
You can just take a subset with low CER, i.e. CER < 5%
Most likely it will be free of noise
Also note that some dataset parts, i.e. prank calls, are very noisy by definition

(3)
Hard to say, but probably radio_2 or youtube

[snakers4]

We do not have proper labels for speakers in this dataset

Though we are planning to apply our speaker encoders retroactively, as soon as we have trained them

[JBloodless]

Ok, what speaker_set in public meta means? Can I believe that speakers in one set is the same (have same pitches, spectral coloration etc.)?

[snakers4]

Ok, what speaker_set in public meta means? Can I believe that speakers in one set is the same (have same pitches, spectral coloration etc.)?

It is the same set of speakers within the same speaker_set
You can believe that if speaker_set_1 != speaker_set2, then in any 2 given audios from set1 and set2 speakers will be 100% different

[snakers4]

I.e. speaker_set1 may contain N speakers
speaker_set2 may contain M speakers different from N