With a letter-level Markov chain. It is trained on real names from a tradition and learns which letter tends to follow each group of letters. To invent a name, it starts at the beginning and keeps drawing letters with those probabilities until “end of name” comes up. Click any name to see its path step by step.
The order is how many letters the chain looks at to decide the next one. With 1 you get unpronounceable names; with 4 it almost only repeats names that already exist (it has learnt the corpus by heart: overfitting). The “copied” counter measures it on every batch.
Choose My list and paste at least 10 names (one per line): those of your town, of a saga, of your characters. The chain will learn how they sound and invent more in the same style. Nothing leaves your browser.