you could use a noise source and sample its sign. Because the bits should not be correlated between them, the bandwidth of the noise has to be large compared with the sampling frequency (so the correlation time is shorter than the sampling period).
But... why not a pseudo random generator based in linear feedback shift registers? It is easy, practical, and performance guaranteed!
Why is a PN generator not good enough? You can use a large sequence, so the lowest frequency will be really low.
Other possibilities are using so called 'gold sequences' or (better but much more complicated) fractal generators. All these circuits can be built with a single CPLD.