Hi! I saw that during preprocessing you split examples of vocalset into 3-second chunks and treated them as individual samples when calculating accuracy. I'm wondering why the metrics are not based on gathered predictions of different splits from the same track just like how you did with Giantsteps if that's the best practice. Thanks!
Hi! I saw that during preprocessing you split examples of vocalset into 3-second chunks and treated them as individual samples when calculating accuracy. I'm wondering why the metrics are not based on gathered predictions of different splits from the same track just like how you did with Giantsteps if that's the best practice. Thanks!