Datasets:
Speaker ID Confusion in the Dataset?
I've downloaded your VoxVietnam dataset, and after extracting the Parquet files and saving them as audio files, then mapping them to individual IDs, I discovered an issue when checking the test set. I found that many IDs contain more than one voice. A prime example is id00005 from the test set.
Could there be an error in the data processing? Or is this an issue with the file upload to Hugging Face? I'm worried that many IDs in both the test and training sets might have similar problems.
Hi Tri,
We highly recommend using the VoxVietnam-O in the README as the test set. Due to the complexity of public YouTube content and how the pipeline's algorithm works automatically, label errors are quite inevitable, we just try our best to minimise them.
Hi lggvu,
Does the same thing happen with training?
Hi Tri,
The confusion in automatic pipelines is hard to avoid. Our team can only manually check the test set. For the training set, you might want to further carry out post-processing steps if you don't mind reducing the data size, even though it was empirically better to train on VoxVietnam compared to the old Vietnam-Celeb.