That is definitely the most paranoid approach, the device guides the user to generate the words manually, then user input the 11 words in the device and device calculates the 12 word.
The combining approach I was more thinking about the non expert user that might use for example a non professional dice and might end-up with a low entropy. But it can probably be explained.
Coin flips probably provides a better randomness, and you can use multiple different coins that will best. Besides everybody has coins.