Teaching a Computer to Speak Whale

Deep in the ocean, sperm whales click at each other in patterns that sound almost mechanical, rapid bursts researchers call codas. For decades, these clicks were treated as noise. Now, a large international research project is trying something far more ambitious: teaching artificial intelligence to find structure, maybe even meaning, inside them.

Globoko v oceanu si kiti glavači klikajo v vzorcih, ki zvenijo skoraj strojno, v hitrih rafalih, ki jim raziskovalci pravijo kode. Desetletja so te klike imeli za šum. Zdaj pa velik mednarodni raziskovalni projekt poskuša nekaj precej bolj drznega: umetno inteligenco uči, naj v njih najde strukturo, morda celo pomen.

The team records thousands of hours of whale clicks using underwater microphones and robotic tags, then feeds the data into machine learning models originally built to translate unrelated human languages without a dictionary. Early results suggest sperm whale codas may combine small building blocks the way human words combine sounds, and some clicks appear to shift subtly depending on context, almost like tone of voice.

Ekipa s podvodnimi mikrofoni in robotskimi oznakami posname tisoče ur kitovskih klikov, nato pa podatke poda modelom strojnega učenja, ki so bili prvotno narejeni za prevajanje nesorodnih človeških jezikov brez slovarja. Prvi rezultati kažejo, da so kode glavačev morda sestavljene iz majhnih gradnikov, podobno kot so človeške besede sestavljene iz glasov, in da se nekateri kliki rahlo spreminjajo glede na okoliščine, skoraj kot ton glasu.

No one is claiming a full whale dictionary yet. But for the first time, researchers have a real, testable hypothesis: that somewhere inside those clicks sits a genuine structure, waiting for the right method to unlock it.

Nihče še ne trdi, da ima cel kitovski slovar. Vendar pa imajo raziskovalci prvič v rokah resnično hipotezo, ki jo je mogoče preveriti: da nekje v teh klikih tiči prava struktura, ki čaka na pravo metodo, da jo odklene.