It seems to be a non physiological model. It uses FFT for frequency resolution, but ear does not perform any FFT. Most recent auditory perception models use gammatone filters to model the frequency resolution of the tympanic membrane in cochlea.
Nevertheless, what ever model you use it does not solve the problem of 'flat' completelely. As also is the case with this model, it requires input signal from where the loudness is calculated. What is your universal standard input signal? Standard music? 🙄
- Elias
Nevertheless, what ever model you use it does not solve the problem of 'flat' completelely. As also is the case with this model, it requires input signal from where the loudness is calculated. What is your universal standard input signal? Standard music? 🙄
- Elias
Good catch Dan.
That falls in nicely with the Kates and Salmi papers: "six FFTs are calculated in parallel, using longer signal segments for low frequencies and shorter segments for higher frequencies" The author is looking more for a model to simulate perceived loudness for any arbitrary waveform, but if you have perceived loudness, then frequency response is just loudness vs. frequency.
Once again, the steady state measurements are misleading.
David S.
Elias, why roll your eyes? Just yesterday you wanted to measure bass using a 10 dB range over a 50 ms window claiming instruments had no steady state. Oye. Your no guru. Interesting that the real gurus disagree with your assessment and don't roll their eyes. All of them. I don't have any problem with flat now(and no, not steady state). No one really enjoys it when people behave disrespectfully and sarcastic. It just makes me want to fire back in the same manner as I have in the past. Then where does the conversation go? Jerry Springer(look it up)? If that's your thing, I ask that you just don't address me or reply to me as I don't want to engage in abusive conversations at all. Thank you.
Thanks Dave. I did a lot of searching last night, but only came up with a few good leads. It's certainly an interesting topic, but it seems very little research has been done on the exact problem that I can find. I guess I'll need to accept a very large grey transition range or come up with my own. Maybe that paper would be worth a purchase--it looks to make things even more complicated and probably more accurate. One thing is certain, everything points in the same direction, but to me I'd like to have a bit clearer definition of the zones. I need to find the Kates and Salmi papers.
Dan
Thanks Dave. I did a lot of searching last night, but only came up with a few good leads. It's certainly an interesting topic, but it seems very little research has been done on the exact problem that I can find. I guess I'll need to accept a very large grey transition range or come up with my own. Maybe that paper would be worth a purchase--it looks to make things even more complicated and probably more accurate. One thing is certain, everything points in the same direction, but to me I'd like to have a bit clearer definition of the zones. I need to find the Kates and Salmi papers.
Dan
Let's leave personal remarks aside and get on with the subject.
Here's what I do to set the position of the forward lobe in crossover design, to "align" the drivers:
There's a video in that thread that shows the process.
The basic idea is to choose the crossover frequency where the beamwidth of the horn and the direct radiating midwoofer are matched in the horizontal plane. Then set the phase relationship between the two drivers that centers the forward lobe in the vertical plane. I do this by looking at the vertical nulls above and below the speaker. The phase relationship is set by acoustic and electrical crossover slopes, by all the electrical, mechanical and acoustic reactance in the system.
The reversed polarity null thing was first proposed by Altec:
More on the whole design process here:
There's a video in that thread that shows the process.
The basic idea is to choose the crossover frequency where the beamwidth of the horn and the direct radiating midwoofer are matched in the horizontal plane. Then set the phase relationship between the two drivers that centers the forward lobe in the vertical plane. I do this by looking at the vertical nulls above and below the speaker. The phase relationship is set by acoustic and electrical crossover slopes, by all the electrical, mechanical and acoustic reactance in the system.
The reversed polarity null thing was first proposed by Altec:
- Altec Application Note 9, "Polarity and Phase"
More on the whole design process here:
To everyone interested in determining the 'flat':
Evaluation of Different Loudness Models with Music and Speech Material, AES 117th Convention 2004
http://www.tcelectronic.com/media/skovenborg_2004_loudness_m.pdf
- Elias
Evaluation of Different Loudness Models with Music and Speech Material, AES 117th Convention 2004
http://www.tcelectronic.com/media/skovenborg_2004_loudness_m.pdf
- Elias
to dantheman:
If you don't like my plots you can project your criticism against the work of prof Eberhard Zwicker, well regarded scientist and the inventor of the Bark scale. What I present is a measured room impulse response in time-freq resolution of the Bark scale. I could guess you never heard about him and are not familiar about the subject? It's hard to argue with someone who don't have the basic knowledge. If you understand or not is not very interesting to me as I have no determination to convince you in this matter or other matters. I just provide.
- Elias
If you don't like my plots you can project your criticism against the work of prof Eberhard Zwicker, well regarded scientist and the inventor of the Bark scale. What I present is a measured room impulse response in time-freq resolution of the Bark scale. I could guess you never heard about him and are not familiar about the subject? It's hard to argue with someone who don't have the basic knowledge. If you understand or not is not very interesting to me as I have no determination to convince you in this matter or other matters. I just provide.
- Elias
Several points here:
1) Dave is posting from Toronto (I presume) at 7 am on a Sunday morning. Wow - thats dedication, or serious insomnia. I'm just getting up and to my computer at about 1 PM (late night party at Dave Clarks with Tom Nousane, Arny Krueger and the regular array of audio fanantics.)
2) We need to be clear on the situation when we talk about different measurement techniques. What one uses to get a speaker only response is very different than what one would want to use to get a room/speaker response. This needs to be clear in the posts.
3) For room/speaker it is clear from many referenced sources that the integration time of the measurement (window time if you will) must be larger at LFs and shorter at HF. Its not clear precsiely what those times should be but there also isn't a huge range of opinions either(I would use 10 ms above 1 kHz stretching to 100 ms at LFs). To me this seems somewhat easy to perform, but admittedly I have not done it. Although, what I typically do is not that much different. I always resample the frequency data to convert from constant bandwidth to constant % bandwidth (the ear is clearly closer to constant % bandwidth). When plotting this information I find about 200 frequency points is all that can be resolved very well so I generally use this number. With only 200 points it would not seem to be too hard to do a different window length for each one of these points (or better yet just those points below 1 kHz, which is about half of them). Sure thats 100 different FFTs on a large number of sample points (I typically use 2048 or 4096), but in FORTRAN, for example, this would not take much more time than about 50 secs or less (much less on a faster multi-CPU computer due to parrallelism). I think that this would be worth a try. Maybe I will someday, but at the moment I am out of bandwidth.
1) Dave is posting from Toronto (I presume) at 7 am on a Sunday morning. Wow - thats dedication, or serious insomnia. I'm just getting up and to my computer at about 1 PM (late night party at Dave Clarks with Tom Nousane, Arny Krueger and the regular array of audio fanantics.)
2) We need to be clear on the situation when we talk about different measurement techniques. What one uses to get a speaker only response is very different than what one would want to use to get a room/speaker response. This needs to be clear in the posts.
3) For room/speaker it is clear from many referenced sources that the integration time of the measurement (window time if you will) must be larger at LFs and shorter at HF. Its not clear precsiely what those times should be but there also isn't a huge range of opinions either(I would use 10 ms above 1 kHz stretching to 100 ms at LFs). To me this seems somewhat easy to perform, but admittedly I have not done it. Although, what I typically do is not that much different. I always resample the frequency data to convert from constant bandwidth to constant % bandwidth (the ear is clearly closer to constant % bandwidth). When plotting this information I find about 200 frequency points is all that can be resolved very well so I generally use this number. With only 200 points it would not seem to be too hard to do a different window length for each one of these points (or better yet just those points below 1 kHz, which is about half of them). Sure thats 100 different FFTs on a large number of sample points (I typically use 2048 or 4096), but in FORTRAN, for example, this would not take much more time than about 50 secs or less (much less on a faster multi-CPU computer due to parrallelism). I think that this would be worth a try. Maybe I will someday, but at the moment I am out of bandwidth.
Last edited:
I am trying to figure out where you are going with the appearant loudness algorithms (weightings). Surely you are not suggesting that all regions of the spectrum should be equally loud?
Hopefully, that is not where you guys are headed.
They used C with Leq, including time domain processing.
Interesting also:
"Our initial evaluation of loudness models indicated that
none of the available models were able to provide an
accurate estimate of the perceived loudness of music
and speech segments. Moreover, none of the existing
loudness models were originally designed to estimate
the loudness of non-stationary sounds such as music"
- Elias
Loudness, as used in that paper, is a single number that is applied broadband. That is not what we are looking for. We need something that can be applied across a spectrum, not a single number.
Could you elaborate on the distinction between "applied broadband" and "applied across a spectrum"?
Regarding a "time" component in the perception of loudness . . . that applies to the original as well as the reproduction. Isn't the point to avoid having the reproduction system (room included) adding temporal distortion in general, not just to avoid mis-perception of "loudness" (which would be a pretty extreme example of resonance/distortion)?
I've been experimenting to implement this:
Analysis of Room Responses, Motivated by Auditory Perception, Lokki, Karjalainen, 2002
- Elias
Analysis of Room Responses, Motivated by Auditory Perception, Lokki, Karjalainen, 2002
An externally hosted image should be here but it was not working when we last tested it.
- Elias
Doug Plumb who is Acoustisoft's developper introduced an interesting feature in his R+D software, he called it "PsychoAcoustic Response" :
it uses a varying gate time. This single curve is composed from measurements that have the gate time change dependent on frequency. Long gate times are used at lower frequencies. The gate time is reduced as frequency is increased until the gate time becomes 5 ms at approximately 200 Hz. At frequencies above 200 Hz the data is smoothed to provide a 1/3 octave resolution. This gives a better indication of overall frequency response as heard be a listener.
Note that those parameters (5ms, 1/3 oct,...) are presets that can be tuned.
The whole manual is really worth reading : http://www.etfacoustic.com/software/RPlusD/RPlusDManual.PDF
Simply windowing the impulse response is not the correct way! The ear's temporal integrator itself has non symmetrical temporal properties. Here from Lokki, Karjalainen according to Plack, Oxenham:
- Elias
An externally hosted image should be here but it was not working when we last tested it.
- Elias
One of my favorite quotes from John Broskie. He's talking about expensive preamps that don't have any tone controls:
“all non-adjustable stereo systems are like broken watches: only occasionally are they accurate.”
“all non-adjustable stereo systems are like broken watches: only occasionally are they accurate.”
Right time zone, wrong city. I've been in New York doing a small theater install at the Central Park Zoo. For the weekend my wife came up and we went to Carnegie Hall! (Toronto Symphony and Itzhah Perlman!) Probably still a little wound up, so I was tapping away at the computer. Sounds like you had a good party. I've met Dave Clark several times but don't know him very well. Seems like a sensible guy.
Now I consider this a bit of a breakthrough. Lets think about the implications. What if, just what if, the energy outside the integrating window (therefore the energy that must be discarded in the measurement) has a different spectral balance than the energy within the time window? Clearly the windowed response curve will have to be different than the unwindowed (steady state) curve. If the speaker's directivity climbs with frequency (very common) then the later sound will be duller than the earlier sound. Room absorption is equally likely to climb with frequency. In large rooms even HF air absorption is a factor. The bigger the room the more likely that the steady state response rolls off in the highs, yet the early sound, falling within our ever shorter window, will be considerably flatter.
I agree there are a number of procedural questions, such as what time vs. frequency profile, what window shape (a few have used exponential windows) and should the window gain be constant (unity at t = 0)?
Flat power response achieved by tilting up the axial response sounds too bright? The model explains that. Cinema X curve needs more rolloff as the number of seats increases? The model explains that.
David S.
What we want to know is what is the spectrum of sound in a room with those reflections that are integrated in with the direct sound. This is not a single number, but a level versus frequency type of curve where the level is adjusted to be more psychoacoustically correct. The loudness in the referenced paper was a single number for a broadband compex signal - its not applicable to the problem as I stated it.
What Doug Plumb is doing is a start, but 5 ms above 200 Hz is not right, and neither is 1/3 octave. But it is important to note that the smoothing and the time window are seperate things. In reality one needs a varrying time window AND a varying smoothing function.
- Status
- Not open for further replies.