I don't believe this could be true, and the linked article doesn't have the word "mix" anywhere. I imagine that there is no mixing happening until after decryption on the client device. Of course this means that every audio source goes to each recipient discretely, which means more bandwidth, but audio (especially near-silent moments therein) is lightweight enough for reasonably sized groups. Obviously this same n^2 scaling issue happens with the video anyway which is never mixed.