Comment by nickpsecurity

The page says it was trained on under 100 hours of audio. Then, the link says “we employ large pre-trained SLMs, such as WavLM, as discriminators with our novel differentiable duration modeling for end-to-end training.” I don’t have time to read the paper to see what that means.

Depending on what that means, it might be more accurate to say it was trained on 100 hours of audio and with the aid of another, pre-trained model. The reader who thinks “only 100 hours?!” will know to look at the pretraining requirements of the other model, too.