Kikiwake is a small, fully local language identification model built on a few layers of Embedding Gemma 2's audio encoder. It identifies a spoken language from about 1.5 seconds of speech without transcribing it.