CSPRNGs require random seeds. They're the obvious right thing when you can generate 256 (or whatever your security level is) bits that are truly unpredictable, independent, and of equal probability 0 and 1, but when generating significantly more is hard. If you can't generate 256 truly unpredictable bits, the CSPRNG doesn't help you much. And if you can generate an unbounded number of them, it's not clear that there's a downside to just using the TRNG data (but there's certainly the argument that it's better-studied and more robust to use a CSPRNG anyway).
Whether this particular method of random bit generation is in fact sound is a different question entirely (and worth asking of all ways to generate those seeds, including the Linux kernel's entropy divination code).
I know most of you take this stuff seriously in your codes and rely on the well know cryptographically secure random number generators: https://en.wikipedia.org/wiki/Cryptographically_secure_pseud...