Electrical noise (including RF noise) is really random, as in it is impossible to predict exact value.
It does have non-flat spectrum, meaning some values are more probable than others, but that only means you need to whiten it. (A rough analogy might be a 6-sided die labeled with 1,1,1,2,3,4 - yes, number 1 is much more likely to come out. No, this does not make it "not really random", and some trivial math can produce ideal random stream out of it)
The only problem with audio input is that you may end up with non-random value - like all-zero output. But properly implemented whitener should detect this and stop outputting any value at all.
it's an often-made mistake where random generation / randomness is confused with probability distribution. Having said that, I don't know (as is in really don't know) if RF noise is unbiased; doesn't sound like it?
If you are talking about DC bias (as in, long term average of raw readings), then "unconnected audio input" is pretty likely to have it - it's easy to introduce via component tolerances, and there is no real reason to keep it exactly zero for audio purposes. But it's also pretty trivial to fix in software.
If you are talking bias in more general sense, then audio input noise is non-uniform in the frequency space, for example there is low-pass filter which filters out high input frequency, and it will affect noise values too. Good whitening algorithm is essential.
The good news however is there are many noise sources which are actually caused by quantum effects in electronic parts, and therefore completely unpredictable. Even if NSA recorded all RF noise, they still could not predict what the ADC will capture. (But they might be able to capture digital bits as they travel over the bus...)
If we were dealing with pure cosmic background radiation, or inside a Faraday cage, sure.
What I'm referring to are things like radio broadcasts, 60 Hz hum from power lines, noise put out by switching power supplies, and that sort of thing.
Just having a bias, as in your example, would be still truly random. If you knew that every tenth roll you'd get a 3, it would no longer be random. When your random number generator can be influenced by the outside world, it's no longer suitable for cryptographic use.
It does have non-flat spectrum, meaning some values are more probable than others, but that only means you need to whiten it. (A rough analogy might be a 6-sided die labeled with 1,1,1,2,3,4 - yes, number 1 is much more likely to come out. No, this does not make it "not really random", and some trivial math can produce ideal random stream out of it)
The only problem with audio input is that you may end up with non-random value - like all-zero output. But properly implemented whitener should detect this and stop outputting any value at all.