[Audacity-devel] 24-bit audio and plot spectrum

[Doc N. <doc...@po...>]

Martyn Shaw wrote:
> Sorry, this is only half a post, but I am a bit inundated right now...
>
> Doc Nelson wrote:
>> During testing for 24 bit operation, I encountered a Plot Spectrum 
>> with a purple screen, the usual -90dB at bottom but -102 at the top.
>>
>> To determine what triggers this, I made a tone of 4410 Hz with 0dB 
>> amplitude (using Generate with amplitude of 1 makes a signal at 
>> -.4dB, so an amplify of .4 pushes it to full amplitude. Close 
>> inspection did not reveal any clipping.
>
> This is bogus.  You made a signal with 10 samples per cycle, and a 
> sinewave, so the first maximum occurs at 360/4 = 90 degrees but that 
> doesn't correspond with a sample.  The samples are at 360/10 degrees 
> so the 1st max ones are at 72 and 108 degrees. 
> 20*log10(sin(72degrees)) = -0.436 dB.  If you were to upsample (and so 
> fill in the gaps) (Tracks -> Resample... -> 96000 (eg)) then there 
> would be values exceeding 1.  Recent changes have meant that they 
> don't get clipped internally, but you may not have that version.
OK, you're right. I was a bit lazy and wanted an integer divison, so 10 
was conveniently done in head. I'll run the experiment again dividing by 
8 or 12. I'll set the sampling rate to 48k to make arithmentic simple
>
>> Amplify -96dB does not work -- it produces a signal of about -48. 
>
> We restrict 'Amplify' to a maximum of 50dB per application (I don't 
> know why).
No big deal. It's probably only for rare testing like this that anybody 
would want to make attenuations greater than 50dB. Still, the why is a 
good question. Why not just let it go to -144dB??
>
> I'm going to have to get back to you on the rest...
Thanks for the help. I have learned a lot about Audacity and the team's 
good works.

I would really like to do a bit of code development, but have to confess 
I am intimidated by the complexity of the OS environments and tools and 
methods. Would it be useful to document any of this for a tutorial?

I had a friend/professional software developer do the ALSA build that 
you tutored in the past. He has since moved away and I have lost the 
recipe.
>
> TTFN
> Martyn
>
>> Accordingly, I applied Amplify -48 twice to get a test tone at -96dB. 
>> At that level Plot Spectrum (4096) works as usual showing a peak 
>> signal at 4410Hz with amplitude of -86.7. The vertical scale of the 
>> plot differs significantly from the measured values from the cursor. 
>> This result is solidly repeatable -- no variations over several tests.
>>
>> I next applied Amplify -6dB repeatedly until the Plot Spectrum screen 
>> went purple. At a level of -114dB the Plot Spectrum still operates as 
>> usual, but there is no signal displayed. At -117 the purple screen 
>> appears.
>>
>> I also made some measurements of this test tone at 0dB, -24dB, -48dB, 
>> -72dB and -96dB for FFT widths from 128 to 4096 samples.
>>
>> In all cases, the level indication of the Output Playback Meter 
>> precisely matches the sinusoidal peak amplitude so there is no 
>> attempt to apply square roots of 2 for RMS or fudge factors to 
>> convert to dBU. The output meter scale is dBV peak.
>>
>> Cursor-measured spectrum peaks tracked the 24dB attenuations of the 
>> signal. At each level, I measured the peak at several FFT widths and 
>> find the the cursor-measured peak drops by approximately 3dB for each 
>> halving of the FFT width from 4096 down to 128 samples. But that 
>> number varies quite predictably as follows 3.7dB, 2.8dB, 3dB, 2.6dB, 
>> and 3.5dB as we go from 128 to 256 to 512 to 1024 to 2048 to 4096 
>> samples at all signal levels.
>>
>> Apparently the tool scales results by the square root of the 
>> bandwidth producing results in dB/rootHz. My HP audio spectrum 
>> analyzer has that feature as an option. We might want to make that 
>> correction optional by a setting in Preferences/Spectrograms.
>>
>> Suggestions:
>> I suggest that the Preferences/Spectrograms wants to have a feature 
>> like Preferences/Interface so users can choose to have the spectrum 
>> vertical scale go down below the noise floor of their 24-bit sampler. 
>> There is some logic to having the spectrum analyzer inherit the 
>> selection from Preferences/Interface. It also seems as if we ought to 
>> have the vertical scale of the graph match the calibration of the 
>> cursor measurements. The differences are in many cases quite large. I 
>> did not attempt to quantify once it was clear that the cursor was 
>> entirely reliable.
>>
>> Questions:
>> I would very much like to know what the Plot Spectrum does when we 
>> select a patch of signal longer the the 2^N FFT width. Does it take 
>> the 2^N samples from the left edge, right edge, average several 
>> spectra across entire window or what?
>>
>> Are we using the FFTW freeware from MIT?   http://www.fftw.org/
>>
>> Thanks
>>
>
>


-- 
Gary Nelson Porpoise Research Institute
Port Townsend, WA, 98368


-- 
Gary Nelson 
Porpoise Research Institute
Port Townsend, WA, 98368