[Help]: Metis: A Foundation Speech Generation Model with Masked Generative Pre-training

[xczhusuda]

Hello, thank you for sharing your work.
After reading your paper, we have a few questions we’d like to ask.

Based on our inspection, the Emilia dataset contains incorrect speaker labels. In some cases, according to the speaker logs used for labeling, a conversation might have labels in the sequence speaker1, speaker2, speaker1, but the second instance of speaker1 is sometimes assigned as a third distinct speaker.

However, we found that your target speaker extraction results are very good. Did you perform any additional cleaning of the data? If so, what method did you use? Or did you train directly on the full dataset without extra cleaning?

We look forward to your reply.