we worked with spectrograms. the spatial prior is useful for audio and i was starting from this (very good) baseline model from bytedance:

→ View original post on X — @jxmnop