Don't recall "word"; a true word is always at least 32 bits, but heretics from x86 world spoiled it finally. Also, all such "octabyte" are long and ugly. Imagine "octaoctet", huh? No, the true words shall sound quite different, to avoid any chance to mix them in the same context.
For 16 bits, the solution is obvious: hexa- -> a sheck. Yes, some like "shack", because 16 bits are now enough only for an old hut of grass and branches.
(Russian: 1 ше́ка)
For higher values, it's some harder. We need an inspiration source not tied with traditional IT languages and nevertheless usable. After some search, I've stopped on Crimean. So, 32 bits = tuze (from: 32 = otuz iki) (Russian: 1 тью́за)
64 = amada, because of: altmış + dört; OTOH the consonant skeleton AMD hints at AMD64 invention, the most popular 64-bit computing source now. Vowels are arbitrary (but, "umudu" is definitely the worst). (Russian: 1 ама́да)
128 = yermezeck (yüzyirmi sekiz); yep, it's long, the same as the data value :)
1 sheck = 2 bytes (octets, pardon my French) = 16 bits
1 tuze = 2 shecks = 4 octets = 32 bits
1 amada = 2 tuzes = 4 shecks = 8 octets = 64 bits
1 yermezeck = 2 amadas = 4 tuzes = 8 shecks = 16 octets (a minimal useful cache line size) = 128 bits
Enjoy and use it thoroughly.
We also shall get rid of awful IEC kibi/mebi/gibi (who, the hell, can distinguish kibi from gibi in a noisy conference hall?) Suggestions:
* 1024^1 = kano-
* 1024^2 = maro-
* 1024^3 = gaido-
* 1024^4 = testo-
* 1024^5 = pempo-
* 1024^6 = erfo-
* 1024^7 = zelmo-
* 1024^8 = yorgo-
Example: 2^31 bits = 2 gaidobits = 64 marotuzes.