Good. We have processed DC offset issue with tmuikku many days ago on Finnish forum and additional hassling here does not look very helpful. Features, philosophy and using of VCAD Merger tool should be discussed somewhere else.
ARTA view on some measurement problems
After several similar discussions I will add some guidelines for using ARTA in audio measurements.
First, I have to say that many guys in this forum tend to assume that everything can be calculated, which is far from reality.
ARTA philosophy is that everything can be measured, and only widely accepted standardized methods may be used in measurement to alter the measured data. Further analysis is in user responsibility.
ARTA saves measured PIR values without mic FR calibration data but can save FR magnitude with applied mic. compensation. It is up to user to apply measured data in some CAD system (VCAD, VituixCAD etc.) to adjust it for specific analysis/application.
The DC removal is an issues. It you take full measured sequence usually there is no DC value, but nonlinearity of measured system (i.e. clipping) can induce high level of low-frequency components. You can easily see it in estimation of step response. Other source of „false“ DC values is gating. You should be always aware of that, and ask yourself: is my measurement OK. Why I have large virtual DC shift? Maybe I can tray to measure with smaller excitation level or with less noise present.
Is it good to remove DC from gated response? Well, in my experience there is great chance that DC removal can add more artifacts. I recommend that you always use the gate on a part of IR that is well above the noise and later apply smoothing to obtain FR.
There are various approach to FR smoothing. ARTA uses approach that was introduced by Vanderkoy and MLSSA.
First, all DFT bins are used to interpolate/average response on logarithmic axis, then that response is convoluted by 6th order Butterworth filter, to obtain smoothed FR magnitude. For phase, the similar procedure of interpolation/averaging is applied on unwarped phase data.
This approach gives results that are closely comparable to classic octave-band analysers.
Other types of smoothing can be applied, like smoothing on bark scale, or simple averaging (brickwall filtering).
For decades the 1/3–octave averaging was the best choice for the estimation of a loudspeaker in room tonal balance. It is amazing to see Martin Colloms reports in HiFi News and HiFi choice, where he very succesfully use spatial 1/3-octave averaging to estimate loudspeaker tonal balance.
For design of crossover filters we usually apply more points in system response optimization (i.e. 1/12 or 1/24 octave apart). In that case a simple averaging on densier grid is appropriate. Later, we may apply additional 1/3 octave smoothing on total response to estimate the tonal balance.
The resolution of frequency analysis is also very problematic in the reverberation time estimation. The standard 1/3-octave analysis has some leakage on neigborough bands, but if we use brickwall filters we shall have large error when measuring short reverberation time.
The same is problem in STI estimation where ARTA uses the 8th order filters.
You see, we need creative „acoustical“ thinking to decide on what type of analysis to apply on measured data.
Two cases are discussed frequently:
1) Construction of a frequency response from measured free field bass response and far field tweeter response.
2) Applying microphone compensation, with calculated minimum phase of mic. compensation, to time domain impulse response data.
3) Calculation of minimum phase frequency response for crossover design.
In both cases there will be always some compromise and ARTA can’t predict your needs.
Hopefully cases 2) and 3) can be skipped if you buy better microphone (at least Earthworks M30).
Best regards,
Ivo
After several similar discussions I will add some guidelines for using ARTA in audio measurements.
First, I have to say that many guys in this forum tend to assume that everything can be calculated, which is far from reality.
ARTA philosophy is that everything can be measured, and only widely accepted standardized methods may be used in measurement to alter the measured data. Further analysis is in user responsibility.
ARTA saves measured PIR values without mic FR calibration data but can save FR magnitude with applied mic. compensation. It is up to user to apply measured data in some CAD system (VCAD, VituixCAD etc.) to adjust it for specific analysis/application.
The DC removal is an issues. It you take full measured sequence usually there is no DC value, but nonlinearity of measured system (i.e. clipping) can induce high level of low-frequency components. You can easily see it in estimation of step response. Other source of „false“ DC values is gating. You should be always aware of that, and ask yourself: is my measurement OK. Why I have large virtual DC shift? Maybe I can tray to measure with smaller excitation level or with less noise present.
Is it good to remove DC from gated response? Well, in my experience there is great chance that DC removal can add more artifacts. I recommend that you always use the gate on a part of IR that is well above the noise and later apply smoothing to obtain FR.
There are various approach to FR smoothing. ARTA uses approach that was introduced by Vanderkoy and MLSSA.
First, all DFT bins are used to interpolate/average response on logarithmic axis, then that response is convoluted by 6th order Butterworth filter, to obtain smoothed FR magnitude. For phase, the similar procedure of interpolation/averaging is applied on unwarped phase data.
This approach gives results that are closely comparable to classic octave-band analysers.
Other types of smoothing can be applied, like smoothing on bark scale, or simple averaging (brickwall filtering).
For decades the 1/3–octave averaging was the best choice for the estimation of a loudspeaker in room tonal balance. It is amazing to see Martin Colloms reports in HiFi News and HiFi choice, where he very succesfully use spatial 1/3-octave averaging to estimate loudspeaker tonal balance.
For design of crossover filters we usually apply more points in system response optimization (i.e. 1/12 or 1/24 octave apart). In that case a simple averaging on densier grid is appropriate. Later, we may apply additional 1/3 octave smoothing on total response to estimate the tonal balance.
The resolution of frequency analysis is also very problematic in the reverberation time estimation. The standard 1/3-octave analysis has some leakage on neigborough bands, but if we use brickwall filters we shall have large error when measuring short reverberation time.
The same is problem in STI estimation where ARTA uses the 8th order filters.
You see, we need creative „acoustical“ thinking to decide on what type of analysis to apply on measured data.
Two cases are discussed frequently:
1) Construction of a frequency response from measured free field bass response and far field tweeter response.
2) Applying microphone compensation, with calculated minimum phase of mic. compensation, to time domain impulse response data.
3) Calculation of minimum phase frequency response for crossover design.
In both cases there will be always some compromise and ARTA can’t predict your needs.
Hopefully cases 2) and 3) can be skipped if you buy better microphone (at least Earthworks M30).
Best regards,
Ivo
This is the part where I "stumble" a bit, at least relative to what has been posted (graphically from Arta) with regard to the latest discussion.
Classically with respect to the 1-2 kHz octave you've typically got two centered "columns" between 1 and 2 kHz plus half a "column" on either side of 1 and 2 kHz.
In the picture below, between 1 and 2 kHz, I'm seeing a degree of oscillation between about 1.2 kHz and 1.9 kHz that at least to me does not represent what I know as a classical 1/3rd octave-band representation.
https://www.diyaudio.com/forums/attachments/multi-way/790186d1572000390-arta-arta-response-png
Ah, one of my favorite topics
In my opinion, the main advantage of the Earthworks M30 (and similar "quality" microphones) is that their frequency response stays the same for a long time. This is not always the case with "cheap" microphones, so you never know if the microphone calibration is still valid for "cheap" microphones without recalibrating them.
However, even high-quality microphones do not always have sufficiently flat frequency response (where the meaning of "sufficiently" depends highly on the needs / requirements of the application at hand). I do not like the situation where uncalibrated data is shared without providing the calibration data that is needed to get the final, "meaningful" data. I would therefore prefer if the ARTA PIR files would contain the impulse response data that has been calibrated for the microphone response curve. Alternatively, it would be okay if the data files would contain the raw/uncalibrated data, but also the frequency response microphone (or other sensor) in order to allow proper calibration of the raw data. As it is, the the raw data alone is not very meaningful, because one does not know how to treat this data in order to reconstruct the acoustic signal that was recorded by the microphone.
That said, I know that the "ARTA way" of saving the raw data only is what many (most?) audio testing software packages do, so this seems to be the common thing to do. But in my opinion, this is wrong. Users of the raw data may not have the calibration data available and therefore cannot reconstruct the impulse response. Again, this is just my opinion!
Hi,
I do not see "oscillations". What I see is 1/3-octave smoothed response.
It is not band representation, as in RTA, which you can choose in Graph Setup dialog by setting combo box Frequency Axis-> Type to "Octave bands"
Ivo
Length of Butterworth kernel would be infinite in theory. Do you truncate that somehow for convolution? Have you considered any IR window function with shorter kernel and zeros at the ends to avoid cropping with artefacts (or improving performance)? For example kernel of Sine function for 1 oct. smoothing is 2 octaves, and 1.89 octaves with Tukey 0.75 assuming that smoothing band is within -3 dB points of function. Both remove small ripple quite well.
I'm totally okay with existing smoothing in ARTA, but just interested.
It is impossible to fulfill your request as manufacturer of microphones does not give true frequency (or impulse) response. They give difference from reference microphone.
Ivo
ARTA uses Butterworth filter FR truncated at points where response is -20dB, and uses it as convolution kernel in frequency domain of log-frequency interpolated data.
Ivo
Hi Ivo, does that mean the Butterworth frequency response is used as a kind of weighting of the log spaced FR data (so a weighted average of the FR samples) or is the log spaced FR data treated as a sampled data sequence and filtered to constrain its bandwidth? Arguably one could say that a 1/3 octave filtered response should be representable as 6 PPO samples without information loss, but that doesn't seem to be the case for most fractional octave smoothing implementations I've come across, for which such a decimation would result in aliasing artefacts.
Hi John,
Your first assumption is ok, just we need to sum power response, and do proper scalling for power od PSD mode.
For details, read AES paper:
Vanderkooy, J., Aspects of MLS Measuring Systems, J. Audio Eng. Soc., vol. 42, April 1993.
ARTA does similar procedure.
Best,
Ivo
Most microphone frequency response curves are indeed determined relative to some reference microphone, for which the response is "known". But the curves are just the delta to the raw reference microphone output, but rather the delta relative to the "known" reality (i.e., relative to the compensated reference microphone). Some labs also have direct ways to measure the frequency response (like the EarthWorks "spark method").
In practice it is certainly possible to calibrate an impulse response measurement such that it represents the "true" response as it was before it was captured and processed by the microphone; I guess this is what ARTA does. Of course one needs to be aware that uncertainties of the "microphone calibration curve" will also exist in the result.
Oh, Ok.
That is what *I call "classic octave-band" (either vertical bar chart or the less known stepped chart) - and it's what I would expect (in a smoothed variation) for a given octave representation.
*and I don't think I'm alone on this after having done a search on: "classic octave-band".
Below is an in-room 1/3rd stepped (..in reference to Martin Colloms). Note that it still only has 4 different pressure levels between 1 and 2 kHz. Though very fine in detail, that's not what I see when looking at the Arta graph previously mentioned.
Attachments
Last edited:
That paper (from 1994, a small typo) seems more concerned with distortion effects, I didn't come across a discussion of smoothing in it (or in the 1989 paper with Douglas Rife). My concern with approaches that use such a form of band filtering for fractional octave smoothing is that they amount to implementing low pass filtering of the sampled frequency response data by a smoothing kernel which is close (Butterworth bandpass) or identical (brickwall) to a simple moving average or box smoother. Moving average has a very poor response, with the first sidelobe more than 20% and a slow sidelobe rolloff, so the smoothed output has components that are much higher in bandwidth than the smoothing fraction would imply. Taking a log spaced data set at 96 PPO as an example, smoothing to 1/3 octave implies a low pass filtering of that 96 PPO data set with a 3 PPO cutoff and ideally components that are above 3 PPO would be heavily suppressed. Zero phase low pass filtering in the sampled frequency domain is required and moving average is a very poor way to achieve that.
It is a little odd that standards bodies do not appear to have set out conventions for fractional octave smoothing of frequency domain data, if they have I'd be interested to see that. The old "AES-X70 Smoothing Digitally-derived Frequency Response Data on A Fractional Octave Basis" project did not seem to come to any conclusions.
Sorry John,
I send you wrong paper reference. It should be:
Lipshitz, Scott and Vanderkoy: Increasing the Audio Measurement Capability of FFT Analyzers by Microcomputer Postprocessing", JAES, September, 1985.
Ivo
Hi,
I made an FFT analyser in 1985. It was two channel device. I used it several years but it has never become commercial product. At the same time a single channel system MLSSA become widely accepted in loudspeaker design. It was much cheaper than HP dynamic analyser or B&K Fourier analyzer. Later I have made ARTA software with intention to merge best things from MLSSA and Fourier analyzers.
I have implemented classical octave band presentation in real-time modes in ver. 1.9.2 (SPA, FR1, FR2) but not in Smoothed FR window, where only sweeped band analysis is presented.
Ivo
Very enjoyable read Ivo, thank you. Wish I'd come across it twenty years ago. As an aside, REW's log spaced conversion uses zero phase filtering of the linearly spaced frequency response to limit its bandwidth to 48 PPO then the filtered response is sampled at 96 PPO. Developing a method for fast fractional octave smoothing of linearly spaced frequency data made that a more straightforward option.
"The format of PIR files is changed. Marker and cursor positions are saved in PIR files. They can be loaded on File->Open command if that action is defined in dialog opened with command File->Options."
Does that mean there are some new fields in the header section? Or does the marker / cursor data go to previously existing fields that were not used?
I am asking because changing the format might break PIR import functions in other software. That wouldn't be nice.