SpeakerId missing in the new revision
I hope that this was not intentional, since to separate the data into test and train, it is important to separate it based on the speaker (since there are multiple samples per speaker), at least, for my project.
As the speaker label is noisy, we do not use speaker id during creation of train and test sets. Have you explored the option of using district or pincode to create train and test split?
Since I am trying to predict the district given the audio, using district to create splits will not be possible. I could try using pincode. Could you expand on speaker label being noisy? I'm currently using an older revision with speaker label to train my models.
If that works for you, we can meet and discuss online. Please drop a note at [email protected] .