The production of seed-independent collisions for various versions of murmurhash (e.g., http://emboss.github.io/blog/2012/12/14/breaking-murmur-hash...) motivated siphash. In general, when there's no positive proof of collision bounds, I would assume (un)lucky inputs can collide much more often than usual, even when you change the seed.

The difference between murmurhash and xxhash in that regard is that anyone can download code to generate murmurhash collisions, while code to make xxhash cry is still unknown/secret.

XXH3 is also a lot faster for long strings, and I believe comparable to murmurhash for short inputs.