room analysis software for osx?

[DanDan]

Well I wish SSD well. I am a bit confused though. the iInterface surely does the conversion?
Another audio system which uses the iPlatform (Nanostudio) claims the following:-
Stereo 44.1kHz floating-point signal path (only the final output is reduced to 16 bit)

EDITED-
In any case, it is possible that 16 Bit is probably more than adequate, depending on how the Analysis is done within the program. This from REW author John PM.
Seems I got confused between a FM bug and a long distant recollection, and saw 24 Bit as a solution, when in fact the change to 24 or is it 32?, coincided with the bug fix. My apologies.

EDITED-
Just took a quick look at REW5. It appears to be saying 44.1/16
DOH what have I inadvertently started here!
I emailed with John the REW author. Due to the analysis process used by REW 16 Bit is absolutely plenty. I am not sure how the software deals with input from the now normal 24 Bit hardware but this seems safe. BTW that great Frequency follows cursor feature is not working well on the current OSX. I see a Java update pending so I hope this restores this great asset, unique to REW.

I checked some recent FM3 measurements. My FR graphs peak at typically -2dBFS. Hey, I'm good ;-)
The ETC initial spike is less than this but only by a few dB and perfectly believeable.
So that bug is GONE and FM totally rocks.

Jens, my apologies for being a bit shirty about the focus on a throwaway comment.
It seems there was something to learn here.

DD

[JohnPM]

To explain the resolution issue, it is important to distinguish between capturing the original response to the stimulus (e.g. a sweep) and making use of the subsequent derived impulse response.

Capturing the original signal requires sufficient resolution to encompass its dynamic range. In acoustic measurements that is often relatively small, perhaps 40 to 60dB, so in principle as few as 10 bits would be sufficient. 16 bits is more than enough for capturing the signal unless one is very careless with levels. Once the signal has been captured, it undergoes processing to generate the impulse response. Resolution is improved by the processing gain of the analysis method.

As a common example of this, here is a plot of the spectrum of a 1kHz tone, captured at 48kHz and 16 bits. The signal is at -20 dB FS, so the maximum resolution the signal can have is the 96dB of 16-bit data less the 20dB the peak lies below full scale, or 76dB. However, the noise floor of the spectrum is at about -130dB! (the top of the plot is 0 dB, the bottom -140 dB). A 128k FFT was used to produce the plot, and an FFT has a processing gain of 10*log10(length/2), which for a 128k FFT is 48dB. In essence, it is dividing the frequency content of the input into narrow frequency spans, and each frequency span has only a tiny part of the overall noise of the signal, hence a noise floor well below what we might expect from a 16 bit signal.


A similar process gain applies to the calculation of the impulse response. The longer the signal that was captured, the greater the process gain involved. Below is the impulse response from a loopback measurement of a soundcard captured at 48kHz/16 bits with a 1M sweep. The sweep level was -10dB, so the captured signal had again at best 86dB of SNR. However, the noise floor of the impulse response, visible on the left before the peak, is at about -125 dB FS (the peak is approx -9 dB FS). If a 128k sweep was used the noise floor would be approx 9dB higher (each doubling of sweep length adds 3dB to SNR).


If we want to export this impulse response as a WAV, however, for example to import it into another analysis package, 16 bits is clearly going to be insufficient as everything below -96 dB FS would be set to zero. Preserving the resolution we have added through our process gain requires 24 bits or better.

[Jens Eklund]

Originally Posted by DanDan

Jens, my apologies for being a bit shirty about the focus on a throwaway comment.
It seems there was something to learn here.

DD

No worries!

[Nordenstam]

Originally Posted by DanDan

I checked some recent FM3 measurements. My FR graphs peak at typically -2dBFS. Hey, I'm good ;-)
The ETC initial spike is less than this but only by a few dB and perfectly believeable.

No normalize to zero?

Originally Posted by JohnPM

To explain the resolution issue, it is important to distinguish between capturing the original response to the stimulus (e.g. a sweep) and making use of the subsequent derived impulse response.

..

If we want to export this impulse response as a WAV, however, for example to import it into another analysis package, 16 bits is clearly going to be insufficient as everything below -96 dB FS would be set to zero. Preserving the resolution we have added through our process gain requires 24 bits or better.

Great post!