My problem is with resolution. As you said TACT stated resolution from PWM is 8 bit. Where do they get required extra resolution?
Kijanki, though it may seem counter-intuitive, reducing the quantization noise that results from limited resolution is mathematically the SAME as increasing the resolution, apart from any possible side-effects of the digital filtering processes. So they get the required extra resolution by the oversampling + noise shaping + dither processes that I referred to.
Keep in mind that limitation of resolution to a finite number of bits is mathematically identical to summing a noise component ("quantization noise") into the signal. If I recall correctly, assuming random error distribution (which is pretty much assured by applying proper dither), the rms quantization noise amplitude is equal to the lsb increment (in volts) divided by the square root of 12.
The oversampling allows most of the quantization noise to be shifted to higher frequencies than the audio information, where it can be filtered out with minimal impact on the audio. Dither randomizes the process to eliminate "deterministic errors," as the article indicates.
I believe that the other pwm applications you mentioned, delta-sigma converters, sacd, and dds, do the same thing.
Regards,
-- Al