Recording Impulse Responses for Speaker Virtualization

[Brandon7s]

I have been digging into this idea for months without knowing Impulcifer and was about to write a program on my own to implement it. But then I found the Impulcifer, which has such a good completion and elegant algorithms! It is such an amazing job! Can't say how much I love it!

Has anyone made the silicone mold microphones proposed by Fabian Brinkmann? According to his paper, this yields the least inter-measurement variance.

I made a pair of them following his idea but with different types of silicone. This is what it looks like. Though I don't have the equipment to measure its variance, it does seem quite stable during multiple measurements.

The mics are bought from a Chinese company called "Yinglear", costing less than one dollar each. They claim that it has 72+dB SNR and flat frequency response. I don't know whether it's exactly true, but it does have an SNR greater than 65dB from my experience.

I'd be very interested in getting my hands on some of those microphones and trying this out. Do you have a link to where we might be able to order these microphones? I did some quick googling but I was too short on the specifics to be able to narrow down the search much. The specs you quoted are pretty good!

This thread is too long and I haven't read through it yet. I wonder has anyone experienced the loss of precision for high frequency (10k+) measurement? Plus, I do feel very bright in 10k+ when comparing the simulation to real loudspeakers.

I've definitely found that the precision in the high end, above 10kHz, seems to be all over the place and is wildly inconsistent throughout my measurements. I thought that it might be a limitation of the microphones, but maybe not.

 

[conql]

Do you have a link to where we might be able to order these microphones?

I ordered them from 1688.com, a China-specific version of alibaba.com. Link is here: https://detail.1688.com/offer/622086499867.html

Not sure whether they ship outside of China or not.

 

[musicreo]

Theoretically, doubling the number of averages increases the SNR of the measurement result by 3 dB, and lengthening the sweep signal can do that too.

The problem is that averaging and increasing the length also increases the time where your head must avoid any possible movement.

In my experience, lengthening the sweep signal does improve high-frequency accuracy for measuring headphones with Impulcifer, but I haven't tested that in measuring loudspeakers.

I haven't tested this with impulcifer but with REW and I don't remember to have seen a significant change between sweep times for headphone measurements.
One big problem of the measurement is to get the correct timbre. This talk from David Griesinger does explain the problem. But I have some measurements were the timbre was ok although the measurement was not taken at the eardrum. Still I believe that deeper mic insertions improve the final result.

 

[conql]

I haven't tested this with impulcifer but with REW and I don't remember to have seen a significant change between sweep times for headphone measurements.

Yeah that's why I posted an issue on Github. Normally 1 second sweep signal can yield a pretty accurate result, but with Impulcifer, it takes 10 seconds to get the same accuracy.


Edit:

After posting this, I check my codes again and find that it's actually my bad......
I thought leaving 100ms of blank in the recording was enough for compensating system latency, but actually, it's not. So I accidentally cut a small part at the end of the recording, making the frequency response at 10k+ abnormal.
Impulcifer is good for measuring headphones even with 1 second sweep. So the default 6.15s sweep should be quite sufficient already. A longer sweep is probably not very helpful.

 

[morgin]

@conql

I’m wondering with your moulds with mics are you doing stereo recording or have you tried the 7.1 surround too?

So far with everyone’s help I’ve managed to get stunning 7.1 results and I cannot see it getting better but that’s what I thought about some previous measurements and they were improved upon. Just wondering if you did something that got you your good measurements

Also the moulds of the ears is interesting because if we can make those of our ears we can glue the mics exactly in place to them so there is no movement. And use those for measurements. But maybe the sound travelling through our heads has a impact

 

[conql]

@conql

I’m wondering with your moulds with mics are you doing stereo recording or have you tried the 7.1 surround too?

So far with everyone’s help I’ve managed to get stunning 7.1 results and I cannot see it getting better but that’s what I thought about some previous measurements and they were improved upon. Just wondering if you did something that got you your good measurements

I tried both stereo recording and surround sound. But I don't think I have got very good measurements yet.

Stereo recordings are very convincing if the sound source is not near. However, when the source is within half a meter, the localization becomes less authentic. I am not sure whether it's because of the headphones or measurement errors.

I use stereo speakers to create 7.1 surround sound. It's stunning for someone who hasn't tried this experience before, but after a short time, I find it only plausible, not authentic. It sounds too bright in high frequencies and I need to adjust it manually for a comfortable listening experience.

Currently, I'm trying to compensate headphones without Impulcifer so that I can measure headphones multiple times to choose the best curve to reduce inter-wear variance. But I don't know whether this will actually help... I'm also planning to use gyroscopes in my phone to get a more accurate orientation when measuring.

I'd be very interested if you can talk about how your measurements were improved. I am new to this topic and I cannot think of a better way than to increase speaker volume and upgrade microphones.

By the way, what headphones do you guys use for reproduction? I read from some papers that headphones meeting Free-air Equivalent Coupling can produce better authenticity. But I don't quite understand this concept and have no idea what headphones meet this criterion.

 

[Brandon7s]

I tried both stereo recording and surround sound. But I don't think I have got very good measurements yet.

Stereo recordings are very convincing if the sound source is not near. However, when the source is within half a meter, the localization becomes less authentic. I am not sure whether it's because of the headphones or measurement errors.

There's definitely an increase in difficulty with getting good near-field measurements compared to getting good measurements from a distance of greater than 5 or 6 feet or so.

My own personal hypothesis as to why quality nearfield measurements are more difficult to obtain is due to the closer measurements being less forgiving to variances in the binaural mic placement/orientation and to any small movements made by the wearer during the measurement process. I'm honeslty doubtful that the wearer making small movements has any significant effect on results, though I am going to test this out by using a monopod or something else to rest my head on getting the capture, but I think that mic placement and orientation is the primary factor at play here.

In my experience, the closer the match in how the two mics are placed in relation to each ears' structure, the better the results and more clear the stereo image and localization becomes. It's a bit difficult to really call it "localization" at such near-field distances since we don't actually hear sound as coming from the speakers at that range - not while listening to both speakers at the same time. It's more about getting an authentic phantom center than getting a feel for the speakers' location in a 3D space, I think.

I think longer distant measurements are less susceptible to speaker directionality since the room has much more of an impact on the measurement when one is listening and measuring a source beyond near-field range. It's easier to localize objects when there's more reflections to work with so I think that more distant measurements will always have an advantage in localization over their near-field counterparts.

The biggest improvement I've made to my capture process with Impulcifer has been taking off the silicon housing on my Sound Professional microphones and gluing them to a regular pair of foam earplugs, and then gluing the cable to the earplugs in a way that will prevent the cable from the tilting the mic capsules' when worn. This keeps the face of each microphone consistently facing straight out of the ear. Then I adjust the depth of each insertion so that the mic capsule is the roughly the same distance from an arbitrarily chosen part of each ear, judging distance by feel alone. Not only have my results since making this change been the best I've ever gotten, they are also a lot more consistent from measurement to measurement. They sound quite similar when A/B'd and the main difference is in where the center of the stereo image appears.

Here's a picture of my mic setup:

I use stereo speakers to create 7.1 surround sound. It's stunning for someone who hasn't tried this experience before, but after a short time, I find it only plausible, not authentic. It sounds too bright in high frequencies and I need to adjust it manually for a comfortable listening experience.

The brightness is a tough one to work out. I've had very good results in using EQ matching via an equalizer plugin from FabFilter called Pro-Q 3 in my DAW. This method involves taking two instances of the EQ plugin and placing one before the BRIR processing, which is done by a convolution reverb plugin, and then the second instance of Pro-Q is placed after the BRIR. I then load up the same sweep audio file that Impulcifer uses and places that in a track. I set the post-BRIR Pro-Q instance to match the EQ of the pre-BRIR EQ instance. I then play the audio of the sweep and Pro-Q generates an EQ curve that it thinks will get your BRIR-processed audio closer to the pre-BRIR audio.

Its a bit confusing trying to explain it like that, I suppose. Basically, I use EQ matching to get Pro-Q 3 to tell me how it would apply corrective EQ to the BRIR in order to get something that resembles the raw sweep response. I initially did this just to see what would happen, but I've actually had some pretty good results by using the EQ curves Pro-Q has given me. A picture might explain this better.

This the EQ curve that Pro-Q gives me when I compare my lastest BRIR (#99!) to the raw sweep audio. This is for my Anandas:

The really interesting part is that big chunk that is taken out around the 2500hz mark, stretching from 1000hz to 7000hz. That's a -4.77db scoop, plus another more minor scoop of -2db centered at 4040hz. That entire trough takes out a LOT of the harshness I experience on the uncorrected BRIR. I've dropped depth of that trough from -4.77 to about -3 to -2.5db but it's a very noticeable improvement even when reduced to only -2db.

Here's what I ended up using as my post-BRIR EQ settings. All of the filters in the beginning are manual room correction filters, since I've not been using Impulcifer's room correction recently. You can see I'm still using the same EQ bands that Pro-Q generated for me at the 2500hz mark:

That particular 2500hz trough is always there whenever I run these EQ match comparisons with my BRIRs, too. It's not just this one particular measurement that benefits from that cut. I'd be curious to know if this kind of EQ matching could point out some trouble-spots in your own measurements to help reduce your brightness problem. If you want to give it a try then I'll be happy to run the comparison for you if you'd like, all I'd need is a copy of whatever hesuvi-compatible BRIR you'd like me to try it with.

You can download the Pro-Q 3 demo to try the EQ curve out and to try disabling/modifying the bands to see what happens and it works just fine when inserted as a VST plugin into Equalizer APO's configuration, but it doesn't support side-chaining to get the EQ match for the part of the signal chain that is pre-BRIR. You have to use it within a DAW to use that feature.

By the way, what headphones do you guys use for reproduction? I read from some papers that headphones meeting Free-air Equivalent Coupling can produce better authenticity. But I don't quite understand this concept and have no idea what headphones meet this criterion.

My primary headphone is a Hifiman Ananda, which I love. I've also used Impulcifer with my Beyerdynamic DT1990, DT770 (250ohms), Sennheiser HD6XX, HD58X, Audio Technical R70X, and ATH-M50S. The Ananda's are head-and-shoulders above the rest in realism even when using BRIRs from the same measurement session, so headphone quality certainly does make a difference.

I'm not sure what free-air equivalent coupling would look like on a pair of headphones, but I'm guessing open-back designs are about as close as we currently have just based off of the keywords used in the name.

 

[morgin]

I’m not as knowledgeable as everyone else but my best measurements came from using a decent speaker, my headphones are hd560s, using foam with mic glued on at the end and cutting them so they fit just slightly into the ear canal, making sure the mics are facing out and both in very similar positions. Also to make sure there is 0 movement (that part and positioning both the same are a pain)

I also used a laser level and a print out of a protractor to measure precisely the angles for surround sound then marked where I should be looking with some tape. The speakers I found having them around 6ft away gave me the best results. Head movement made a big difference too so I used a swivel chair and made sure my head was exactly above the base so when turning my ears were rotating exactly on point and not making a big circle.

Also a big factor with my measurement we’re not using full volume I found that just above normal listening levels was right.

With all that I don’t use any balance options because I feel they make a negative impact on what I have atm.

 

[conql]

There's definitely an increase in difficulty with getting good near-field measurements compared to getting good measurements from a distance of greater than 5 or 6 feet or so.

My own personal hypothesis as to why quality nearfield measurements are more difficult to obtain is due to the closer measurements being less forgiving to variances in the binaural mic placement/orientation and to any small movements made by the wearer during the measurement process. I'm honeslty doubtful that the wearer making small movements has any significant effect on results, though I am going to test this out by using a monopod or something else to rest my head on getting the capture, but I think that mic placement and orientation is the primary factor at play here.

In my experience, the closer the match in how the two mics are placed in relation to each ears' structure, the better the results and more clear the stereo image and localization becomes. It's a bit difficult to really call it "localization" at such near-field distances since we don't actually hear sound as coming from the speakers at that range - not while listening to both speakers at the same time. It's more about getting an authentic phantom center than getting a feel for the speakers' location in a 3D space, I think.

I think longer distant measurements are less susceptible to speaker directionality since the room has much more of an impact on the measurement when one is listening and measuring a source beyond near-field range. It's easier to localize objects when there's more reflections to work with so I think that more distant measurements will always have an advantage in localization over their near-field counterparts.

That's what I think too. So I choose to do binaural recordings of my smartphone playing music while I'm turning my head around. And by checking localization errors I can easily tell whether my headphone compensation is accurate enough.

The brightness is a tough one to work out. I've had very good results in using EQ matching via an equalizer plugin from FabFilter called Pro-Q 3 in my DAW. This method involves taking two instances of the EQ plugin and placing one before the BRIR processing, which is done by a convolution reverb plugin, and then the second instance of Pro-Q is placed after the BRIR. I then load up the same sweep audio file that Impulcifer uses and places that in a track. I set the post-BRIR Pro-Q instance to match the EQ of the pre-BRIR EQ instance. I then play the audio of the sweep and Pro-Q generates an EQ curve that it thinks will get your BRIR-processed audio closer to the pre-BRIR audio.

Its a bit confusing trying to explain it like that, I suppose. Basically, I use EQ matching to get Pro-Q 3 to tell me how it would apply corrective EQ to the BRIR in order to get something that resembles the raw sweep response. I initially did this just to see what would happen, but I've actually had some pretty good results by using the EQ curves Pro-Q has given me. A picture might explain this better.

This the EQ curve that Pro-Q gives me when I compare my lastest BRIR (#99!) to the raw sweep audio. This is for my Anandas:

The really interesting part is that big chunk that is taken out around the 2500hz mark, stretching from 1000hz to 7000hz. That's a -4.77db scoop, plus another more minor scoop of -2db centered at 4040hz. That entire trough takes out a LOT of the harshness I experience on the uncorrected BRIR. I've dropped depth of that trough from -4.77 to about -3 to -2.5db but it's a very noticeable improvement even when reduced to only -2db.

Here's what I ended up using as my post-BRIR EQ settings. All of the filters in the beginning are manual room correction filters, since I've not been using Impulcifer's room correction recently. You can see I'm still using the same EQ bands that Pro-Q generated for me at the 2500hz mark:

That particular 2500hz trough is always there whenever I run these EQ match comparisons with my BRIRs, too. It's not just this one particular measurement that benefits from that cut. I'd be curious to know if this kind of EQ matching could point out some trouble-spots in your own measurements to help reduce your brightness problem. If you want to give it a try then I'll be happy to run the comparison for you if you'd like, all I'd need is a copy of whatever hesuvi-compatible BRIR you'd like me to try it with.

It's a bit complicated and I don't quite understand Do you mean using the plugin to see the frequency domain of the BRIR recordings and adjust the filters according to that? What's the difference between this and the APO provided graph? Anyway, thanks for your information.

I also used a laser level and a print out of a protractor to measure precisely the angles for surround sound then marked where I should be looking with some tape. The speakers I found having them around 6ft away gave me the best results. Head movement made a big difference too so I used a swivel chair and made sure my head was exactly above the base so when turning my ears were rotating exactly on point and not making a big circle.

It's indeed a good idea to mark the place to look at. Gyroscopes' error is too large, not accurate enough for measurement.

For me I would say that blocked ear channal measurements sound more bright while measurements with only partial blocked ear channal sound better.

I make a pair of microphones without all the silicone stuffs and directly plug them in my ear canals with tapes mounting. And they provide much better results for binarual recordings reproduction. I am speechless now cause I really have spent a lot of time making silicone moulds and etc... Now it proves they're useless.

The problem is, I have two MDR-V6 headphones, one is with the original earmuffs but has bad seal and a lot of sound leakages, the other is with new earmuffs that provide good seal. The blocked ear canal microphones work fine on the old headphone, though not better than the new microphones. But they are significantly worse when measuring the good seal headphone and reproducing binaural recordings on it. I suspect changing the earmuffs affects headphones transfer characteristics, making them less "Free-air Equivalent", thus the blocked ear canal method will bring significantly more error. I think I've got to do some research on that...

And I notice that open ear canal measurement results in more high frequency variance from taking off headphones and taking on again. I wonder how do you deal with this inter-wear variance? For me it's a clearly audible difference.

@musicreo
I'm very interested in your mic setup. Could you share how you make your partial blocked ear canal mics?

 

[Brandon7s]

It's a bit complicated and I don't quite understand Do you mean using the plugin to see the frequency domain of the BRIR recordings and adjust the filters according to that? What's the difference between this and the APO provided graph? Anyway, thanks for your information.

What I'm doing with the Pro-Q equalizer software is using it to compare the sound of a BRIR from Impulcifer to the sound of the unprocessed sound sweep. Pro-Q then looks at the comparison and automatically places a corrective EQ curve to make the audio post-BRIR more closely resemble the EQ curve of the unprocessed sweep. This makes it obvious to see where the EQ of your BRIR dramatically differs from a perfectly "flat" frequency response (to the computer's ears). It's a bit of a long-way around trying to manually EQ out problematic areas by seeing if the EQ software shows anything that really pops out as being different to the theoretical flat response.

In my own case, the EQ software showed that a big dip between 1kHz and 7kHz would bring the final EQ curve more close to flat, and that ended up being the case to my ears when I tried it out. The screenshots I shared show that portion of the EQ that the software pointed out as needing significant changes to the EQ of my BRIR.

You could do this by making changes to the EQ manually and judging the visual flatness of the APO graph, though attempting that myself yielded poor results. The nice thing about Pro-Q is that it generates EQ corrections that are easy to switch on and off and to adjust manually.

 

[conql]

What I'm doing with the Pro-Q equalizer software is using it to compare the sound of a BRIR from Impulcifer to the sound of the unprocessed sound sweep. Pro-Q then looks at the comparison and automatically places a corrective EQ curve to make the audio post-BRIR more closely resemble the EQ curve of the unprocessed sweep. This makes it obvious to see where the EQ of your BRIR dramatically differs from a perfectly "flat" frequency response (to the computer's ears). It's a bit of a long-way around trying to manually EQ out problematic areas by seeing if the EQ software shows anything that really pops out as being different to the theoretical flat response.

In my own case, the EQ software showed that a big dip between 1kHz and 7kHz would bring the final EQ curve more close to flat, and that ended up being the case to my ears when I tried it out. The screenshots I shared show that portion of the EQ that the software pointed out as needing significant changes to the EQ of my BRIR. You could do this by making changes to the EQ manually and judging the flatness via the APO graph, though attempting that myself yielded poor results. The nice thing about Pro-Q is that it generates EQ corrections that are easy to switch on and off and to adjust manually.

I see it now. I'll definitely give it a try when I finish building a new mic setup and sorting out this measurement issue.

 

[musicreo]

And I notice that open ear canal measurement results in more high frequency variance from taking off headphones and taking on again. I wonder how do you deal with this inter-wear variance? For me it's a clearly audible difference.

This plot shows three times removing the mics and headphone. I don't have big differences as long as I use the same mic pairs.

@musicreo
I'm very interested in your mic setup. Could you share how you make your partial blocked ear canal mics?

According to the paper Sound transmission to and within the human ear canal the mic position is not so important for the measurements. Still they show extreme differences between measurement positions.

Post #515 shows my setup I used for my measurements.
I tried a lot of mountings. The last image in the first row provided me my best measurement. The mounting in the last two images use a silicon tube with a small wire so that the capsule can be bent and hold into position. This was an idea I wanted to test recently but I run into a strange SNR problem with many Primo EM258 capsules. This is the reason why I used a big capsule (last image) for the test.

 

[musicreo]

What I'm doing with the Pro-Q equalizer software is using it to compare the sound of a BRIR from Impulcifer to the sound of the unprocessed sound sweep. Pro-Q then looks at the comparison and automatically places a corrective EQ curve to make the audio post-BRIR more closely resemble the EQ curve of the unprocessed sweep. This makes it obvious to see where the EQ of your BRIR dramatically differs from a perfectly "flat" frequency response (to the computer's ears). It's a bit of a long-way around trying to manually EQ out problematic areas by seeing if the EQ software shows anything that really pops out as being different to the theoretical flat response.

I don't understand the reason to equalize a BRIR to a flat frequency response?

You could do this by making changes to the EQ manually and judging the visual flatness of the APO graph, though attempting that myself yielded poor results.

When you use HeSuVi you must consider that you see always the sum of the channels in the EQ-APO graph.

In the moment I started to equalize any bigger channel differences between all left and right channels that are even present after using Impulcifers channel balance correction. First impression is very good.

 

[Brandon7s]

I don't understand the reason to equalize a BRIR to a flat frequency response?

To help counter any errors in the measurement process or deficiencies in the microphones, mostly. Knowing exactly which frequencies differ from flat, you can get some ideas on what spots of the measurement might need additional correction. For instance, the measurements I make with my current mic setup all sound quite a bit less 'sparkly', hi-fi, and clear when compared to listening to the physical speakers. I'm not sure exactly why that is yet but my best guess is that my mics exhibit more high-end rolloff than I thought.

In any case, running my BRIRs through Pro-Q set to match to a flat EQ will quickly provide an EQ correction curve to help offset that high-end rolloff. I then tweak it to taste, which is a lot easier than trying to quickly come up with a manual correction curve by ear from scratch for each BRIR I use.

 

[musicreo]

I made a graph that explains my thoughts about channel balance problems. I think the parts I marked with audible are problematic. Actually they don't do any harm to the channel localisation but result in a exhausting listening experience over time. With my very open AKG701 and HD555 I did not notice this problem much and ignored it but with my HD600 it became annoying. Equalizing this parts really improved the listening experience with the HD600 a lot. The problem is that there is no way to know if left channels should match right channels or right channels should match left channels or if this is also depending on the frequency range. So I think it is just testing and finding the best way. But certain is that in my case the channel balance correction of Impulcifer did a good job but still left some serious issues.

If I would compensate for a flat response in that plots I guess that I would remove the filter function of my ear.