I bought a røde wireless mic which definitely helped. It gives deepfilternet enough good signal:noise to work reasonably well, but I was hoping there was an even better solution.
Instead, sound engineers spend weeks cleaning up spoken dialog by hand in spectrogram editors. It's honestly astounding the magic they can do, but it's also labor-intensive and therefore expensive. They're literally individually EQ-ing every vowel, splicing consonants from one word to another... it's wild.