Bob Axtell wrote:

>> Roman certainly knew his stuff. I'm sure if he said it worked for
>> him, it did. Likewise I'm sure everyone who's tried it and not got
>> the expected performance also knows what they're doing. Roman
>> obviously has great confidence in his method, so why results haven't
>> been replicated is a mystery

> It would seem to me that by now, this technique would be used everywhere
> commercially. Not an insinuatiion, just an observation. Because if it
> could be made to work, it would put 14-bit DACs into the realm of the
> dinosaur. 

Roman seems to base his tests mostly on a sample rate of 19.5 kHz. (At
least that's how I interpret the graphics on that page.) I think what he
calls his "predictive" encoding might simply be a 2nd order delta sigma
DAC: the prediction of the single pole low pass adds an integration level
to the algorithm, which is what increasing the sigma delta modulator order
also does. I haven't done the math, but it looks to me like that. So he's
using a 19.5 kHz delta sigma DAC of 2nd order, with 1 or "1.5" bits (let's
just say 2 bits, to be able to apply more common math). 

Since with delta sigma DACs the bitrate alone doesn't determine neither the
bandwidth nor the resolution but only the combination of the two, it's the
combination of lowpass and bitrate that determines the bandwidth
resolution. 

It seems his BTc8 constant assumes a lowpass time constant of 8 times the
sample time. That gives a cutoff frequency of 390 Hz (for a sample
frequency of 19.5 kHz). That doesn't sound quite enough for voice
communication. But anyway, this is equivalent to an oversampling by a
factor of 25. With a delta sigma DAC 1st order, that's about 35 dB or some
6 bits resolution; with a delta sigma DAC 2nd order, that's about 55 dB or
around 9 bits.

Using a different filter, say 3.2 kHz, the oversampling is only by a factor
of 3, which gives pretty low resolutions in the 2 to 3 bit range. For a
2-bit DAC, it gets more interesting: the oversampling is now equivalent to
a factor of 12, and with a 2nd order DAC that's 40 dB or 7 bits -- good
enough for voice, as is the bandwidth of 3+ kHz. 

So that seems to be the configuration that might work for simple voice: a
2-bit delta sigma DAC 2nd order, with a sampling rate around 20 kHz and a
lowpass filter around 3.2 kHz. 

I'm not sure this is exactly what Roman is proposing, but it seems to be
close. His design is basically a delta sigma DAC of something between 1 and
2 bits, I think that his prediction algorithm is in essence creating a 2nd
order delta sigma modulation, and he seems to work with a sampling rate of
19.5 kHz. Just the filter cutoff is a bit odd, but I may have simply
misunderstood what he's writing about. And this is of course not a really
sharp filter, so the best (audible) result may not have the theoretical
cutoff frequency.

However, I don't think it would work with the 5+ kSamples/s that you were
trying. I also see only the 19.5 kSamples/s pictures at Roman's page, no
mention of lower sample rates. Given that the 19.5 kHz sample rate is
already a border case with 3 kHz bandwidth and 7 bit resolution, I don't
think that 5 kHz sample rate gives an intelligible result with a 2-bit
delta sigma DAC of 2nd order. 

Gerhard

-- 
http://www.piclist.com PIC/SX FAQ & list archive
View/change your membership options at
http://mailman.mit.edu/mailman/listinfo/piclist