Re: [Audacity-devel] Maximum *Table* size

[Steve t. F. <ste...@gm...>]

On Thu, Aug 25, 2011 at 9:18 PM, Roger Dannenberg <rb...@cs...> wrote:
> On 8/25/11 2:44 PM, Steve the Fiddle wrote:
>> Thanks for the reply Roger,
>>
>> Yes, I used (snd-resamplev) with "s" as the carrier, though I used
>> (integrate) to map the modulation signal rather than adding it to a
>> ramp.
>>
>> "Where do you stop?" - I stopped at a billion samples :=)
>>
>> I take your point about linear interpolation, but where/when does that occur?
> The FMOSC function does (linearly) interpolated lookup on the wavetable.
>>
>> I agree that 100K samples should be plenty for single period
>> waveforms, but what I had in mind as one of the "other situations"
>> where a larger table size might be useful was for multiple period
>> waveforms. Specifically, I've been creating tables using additive
>> synthesis, and 100K samples seems to be a bit too close to the limit
>> for the necessary oversampling. I notice that in the v. 3.05 manual
>> entry for (maketable) it says that "Currently, tables are limited to
>> 1,000,000 samples", which in some cases does seem to give better
>> results. http://www.cs.cmu.edu/~rbd/doc/nyquist/part8.html#index333
> I *did* have the sense that 100K was on the low side, but I couldn't
> think of a need for anything bigger. My paper, Interpolation Error in
> Waveform Table Lookup (http://repository.cmu.edu/compsci/502/), claims
> that you get 96dB signal-to-noise ratio with 32 harmonics in a table of
> size 10402, so in that sense, 100K should be good. On the other hand, if
> you have samples with an attack and multiple periods, I can see 100K
> might be a bit low, and I'd rather correct the code to match the
> documentation than the other way around (I don't know why they don't agree.)
>
> So if I'm going to make a change, is 1M samples (4MB) reasonable? One
> reason for the limit is to alert the user to problems rather than slowly
> allocating and filling virtual memory. Also, I'd really like to
> discourage people from writing algorithms that use tables to process,
> say 20s or even 200s of sound when an alternative like snd-resamplev
> would do a good job without duration limitations.
>
> -Roger
>

The "discouragement" works :=)
I've been reworking my additive synthesis code and found that I can
get good results without aliasing with much smaller wavetables than my
previous code required. In fact I've not found any situation where a
table size greater than 100K is actually required.
One thing that does strike me as peculiar though is that
(build-harmonic) seems to give better results with a table size of
2048 than any other size (larger or smaller).

Steve