I am currently designing a ADC converter, that will match Dave's performance, and solve a number of issues that plague conventional ADC's - notably huge noise floor modulation, poor anti-aliasing filters, and poor noise shaper performance.
I know from the work with Dave that the perception of depth needs noise shapers of astounding accuracy; indeed, Dave ended up with 350 dB performance noise shapers, in order to ensure that small signals are resolved with zero error - from listening tests, this is needed to ensure the brain can perceive depth correctly.
Now I have designed a ADC noise shaper that exceeds 350 dB performance (note these numbers are digital domain performance only, so it is an idealised noise shaper - I am only looking at the THD and noise of the noise shaper only). To test the noise shaper I can run Verilog simulations, capture the data, then do an FFT on the data, and then check the results. Before I did that, I thought it would be a good idea to run a similar simulation with Dave's noise shaper. In this case, I am trying to evaluate whether it can accurately encode very small signals, so I am using a -301 dB sine wave at 6 kHz. If it can resolve a signal at -301 dB, then we can safely say that small signals are accurately encoded, at least in the digital domain.
So here are the results:
So this is the digital domain performance of the Dave noise shaper, and frequency is from DC to 100kHz (0.1 MHz).
The 6 kHz signal is perfectly reconstituted at -301 dB. You can see a flat line at -340 dB, but this is just a FFT issue. The real noise floor at 15 kHz is at -380 dB, which is about 100 trillion times lower noise than conventional high performance noise shapers. Note also the noise at 100 kHz is at -200 dB - that is extraordinary low for a noise shaper, and shows why I need to do little filtering on the analogue side.
-301 db is better than 50 bits accuracy.
Now to write the code for the ADC!
Rob