Listen Inc - Correlating Transient Distortion Measurements w Audibility Enhanced Loose Particles eLP
[Music] hi everybody Dave Lindberg in Hong Kong with another episode of the THD podcast uh this is a part two episode and we have Steve Tammy from listen uh continuing on and he's going to be talking about uh perceptual algorithms for End of Line testing and specifically uh something to uh an enhanced measurement to detect loose particles in your speaker measurement so we're going to get rolling on that in a moment but without delay Alti Association our sponsor we encourage everybody to go check out their website and see what they have to offer it's great networking for people in the audio manufacturing industry so without delay Let's uh get into the discussion so the one other thing I just wanted to talk about um speaking of perceptual measurements is you know periodic Distortion harmonic Distortion is not the only kind of distortion out there there's something that we like to call loose particle Distortion but so in the early days um you know we were catching rubb and Buzz but there were cases where you know dirt magnetic chips um foreign particles whatever you want to call them would fall into the voice coil Gap my favorite expression was crap in the Gap and we wanted to be able to catch that because it's also annoying we actually had one customer who was pretty funny who made pro audio loudspeakers and someone left a hammer inside the the loudspeaker enclosure and the customer said I keep on hearing this rattling sound when they when they got the speaker back they found you know that the ultimate loose particle was this Hammer jumping around when the cabinet vibrated but in any case we wanted to be able to catch um you know even just uh the lead wires to the um uh voice coil sometimes would hit the speaker so how could we catch those so I realized later on that it wasn't just loose particles that people wanted to catch they also wanted to catch um rattling buttons and um Fasteners and you know the the speaker and the test box so if we want to catch these what I'll call transient Distortion or some people call them impulsive Distortion there's all kinds of names um they're not periodic they do get vibrated by the physical movement of the loudspeaker of the diaphragm so you know it's it's vibrating everything around it and if something is loose like a um magnetic chip or glue chip in the voice call or it's rattling a button on a smart speaker and we try to analyze it in the frequency domain it just looks like noise because it's not periodic so the more we average the more averages out so we can't really easily look at these defects in the frequency domain it makes a lot more sense to look at them in the time domain so at the back in 2004 when we were developing this algorithm the only other um I'll call it analysis technique that we were aware of was Crest factor and Crest factor is uh of course a very useful um analysis tool and it's pretty simple to implement just by using what we call a track and highpass filter so again we do a fast sign sweep we offset quote a highpass filter to um remove the uh the stimulus and measure everything else everything else in this case is can be loose particles but it can also be rubb and Buzz and that's because not only will loose particles by the way Crest Factor just briefly to remind people is the peak to RMS ratio so if you look at a sine wave the crest factor is like is 1.414 it's very low but if you look at a transient it's going to be very high if you look at Buzz which kind of looks like a transient it's also going to be high typically um over 10 so what people would do is they'd use um this tracking highpass filter and look at the crest Factor versus frequency and it was a single Distortion metric but it didn't isolate periodic from what I'll call nonperiodic Distortion or uh transients so it kind of lumped them together the other problem is background noise also tends to have a high Crest factor I mean random noise is is typically um 20 DV so um we're also going to get false rejects due to factory background noise um so that's no good and of course the limits are tricky too because they change versus frequency there's none of this normalization being applied to the the loudness so what we did back in 2004 is we said well let's do it a little differently we're still going to use a tracking highpass filter and capture whatever you want to call the the crud the Distortion and the noise but instead of plotting it versus frequency we're going to plot it versus time and take the time envelope now the reason we take the time envelope is we need to look at it not like an oscilloscope but we want to look at it as for example dbspl so that we can set limits and instead of um looking at Crest Factor we're going to count the number of transients uh during that two Second Sweep and basically if we have um loose particles or something rattling we're going to get many transient events and if it's background noise we're only going to get maybe a few I mean if a compressor goes off or someone drops something a Forklift goes by in a production line it's not going to happen 50 times in two seconds where a loose particle or a button rattling it's it's going to be at least 10 uh events in a couple of seconds so we can separate out what we'll call the um transient defects or Distortion coming from the product versus the background noise and we can just set a simple limit where we say if it's greater than three transient events it's probably coming from the the loudspeaker but our algorithm back then was still if you look at the envelope it's still a bit tricky to set limits because again they they do change versus frequency so we wanted to come up with a more robust way of doing it so basically um the first thing is we want to separate out the the loose particle form as we called it from the stimulus and basically we created a very good tracking Notch filter so that you could listen to everything but the stimulus um one of the things that always amazes me is that um we've heard of the psycho acoustic phenomena missing fundamental and even though you even though you Notch out the uh the stimulus if you have some second and third harmonic or some Distortion you still hear some sign sweeping because your human ear is trying to fill it in say hey that should be a sign sweep but anyhow I I digress it it kind of caught me off guard when I first did this and I listened to the loose particle waveform without the stimulus I thought I could still hear some of the stimulus like the filter working yeah yeah I mean there's a lot of uh uh signal processing people put in uh amplif you know base amplifiers and uh rock and roll um but we wanted a easier way to set the limits because that was really the biggest challenge people weren't using the loose particle algorithm because it still was getting kind of tricky to set you know they wanted to set a flat line again and they couldn't so we wanted to improve the uh signal to noise the dynamic range and we also wanted to catch um not just loose particles and as I mentioned buttons but things an automotive which they typically call buz Squeak and Rattle and that is you know door panels vibrating and um you know a car has so many Fasteners so many parts and the last thing you want is your um you know your $100,000 EV uh Tesla making some annoying uh rattling sounds because your speakers are vibrating it so so basically um the algorithm looked like this we we enhance uh the ability to extract artifacts which basically we made a better Notch filter we still use envelope analysis but the big difference is that we tried to come up with a better way to catch the Peaks the transients using something called prominence and I'll explain that in a moment but other than that it's the same algorithm from 200 for so again just real quickly we have a sign sweep so this is just like you saw before but didn't hear before you know the um typical kind sign chir it's actually step sign but the transitions are so smooth from frequency to frequency the phase continuous it sounds like a chirp the next thing is now we captured the response we recorded um in this case a speaker with some really obvious loose particles it's a solder bead inside the voice callil we P we picked a particularly bad one so it would be easy to hear but this is kind of a severe case it is a severe [Music] case hopefully you can faintly hear the rattling okay good but now when we extract the uh we remove the stimulus from this is exactly the same recording without the stimulus the stimulus notched out now we can not only see but we can also hear the loose particles very clearly and I think at the beginning you probably heard the stimulus a little there was some second and third hormon at low frequency so that is the missing fundamental so now we got to figure out how are we going to put limits on this well we really need to come up with a um way to measure the level and we want to measure it ideally in dbspl so we do the RMS and we get the envelope but as you can see that the um Peaks do vary versus uh frequency which is versus time time and frequency in this case so how can we come up with an easier way to set limits because you notice some of these Peaks things kind of bounce around and you don't want to count that twice you don't want to count that twice so what we do is we use um what's called prominence and I'll explain that here now and so instead of um it really comes from cograph um where you're looking at a mountain range and typically we think of elevation you know it's like okay um you know how far above sea level is this peak but if you're hiking a mountain what really matters is what is the difference between where you're starting and where you're ending and that could be called the prominence so in this um elevation elevation C obviously is the highest but if what we really want to know is the difference between the nearest uh Peak and the near nearest null or L we would say what's the prominence and in this case it's actually prominence a which is going to be the biggest difference so that is going to help us be able to detect when there's a spike that is Audible and I'll I'll show you some examples in a moment so here is now the envelope analysis but prominance is applied to it so now it cleans up the envelope quite a bit where it's very obvious that we have some sharp spikes due to these transient events you can see there's some little ones in the background but we've effectively increased our dynamic range and made it easy to set limits with a straight line again and what we do is we just count the number of times the events go above the threshold this is just a snippet of the total sweep but if we looked at the entire sweep we would actually see 110 uh of these spikes that you heard during the sign sweep and again if it was background noise it' probably just be a few so just to kind of um repeat we got the response wave form the recorded response wave form we Notch out the fundamental we get what we call the loose particle waveform which we can play back and listen to and correlate and it's not just going to be um loose particles but you can also hear Distortion which is pretty entertaining as well if you listen to rub and Buzz this way then we take the envelope and then we take the prominence and if I just let this thing cycle you can see very clearly with your eye how the spikes show up and get filtered essentially and we end up with the prominence so then the solution for that is the factory worker notices that and takes a little air gun and blows out the speaker and tries again if they can okay um more often than not it goes people most people we're talking to today have never been to the factory but yeah they're always got the air guns blowing off the speakers before they test them so that's kind of a standard operating procedure yeah and you're right and you know I always tell customers measure the speaker pointing down um there's a couple reasons for that one is you're not going to get stuff falling on top of the speaker like you might in a chamber number two if there is a loose particle it's more likely to bounce around off the um the diaphragm or the uh dust cap so um if you point the speaker up the loose particle might just get stuck in the bottom and you won't catch it and then you ship it and then it bounces around and your customer says hey I can hear some something rattling around um but yeah you're right uh I I even do uh demos of this with h salt and pepper I just put on top of the loudspeaker and I got to blow it off at the end to kind of you know get rid of them right all right so last quick sales pitch here and then we can move on to your questions um but basically yeah we're we're trying to refine things and I'm a big believer in uh you know product Improvement for through refinement and so we're constantly trying to make it easier um improve our algorithms make uh setting limits easier be able to listen to what we're measuring measuring and correlate objective with subjective and um yeah and if it can catch other problems on your product you know not just the uh the transducer I know the automotive loudspeaker guys they they hate it when um one of their speakers gets mounted improperly in a $100,000 car and then the car manufacturer blames the uh the loudspeaker manufacturer and it's like it's not our problem it's it's you did a bad job of fastening that loudspeaker in the in whatever the door panel so it's also good all right so uh yeah we've got a got a few minutes left so uh What uh what's the future holding for listen well good question um you know it's really a bunch of different things that I'm trying to uh work on as usual probably too many things um but I mean I think everyone is uh aware of you know not just smart um speakers but also um things like spatial audio um I personally have always been curious about you know how do we determine a source localization how do we hear it and you know there's anyone who is um been doing this is really aware that there's there's a lot going on um and it's not just you know headphones but it's also cars um and it's really thanks to um you know companies that are trying to work with like AES and Alti um and create uh the source material we're actually trying to figure out ways to uh to measure the source uh location yeah so you're your thoughts like so measuring these new things like people talk about measuring spatial audio or immersive or whatever this this new buzzword is and I'm sure this this is a continuing progress and this is how they come up with new products and New Market segment so we got to live with it but uh I guess what are your thoughts on on that spatial audio Yeah so I I you know I really had a uh I guess you could say an eye opener and an ear opener when uh about two years ago at Christmas we bought our son um a VR headset and you know I thought it was just for gaming and I was not all that excited about it but um you know I put on the VR headset and he had me take the um roller coaster ride with the Via headset on and I was just absolutely blown away that um both by the video and the audio and that without actually um feeling any of the vibration or the the motion my brain was telling me I was going up and down and getting slightly uh you know seasick or whatever you want to call it so I was like wow this this really is pretty amazing and and I wanted to figure out a way to uh to correlate um you know the the spatial audio to ways to measure it I mean I think one of the things that really um convinced me that this was the future was uh just starting to play around with binaural recordings and listening to um ideally you know my own hrtf which is is a it's of course extremely difficult to get your your own personal HT hrtf measure because it takes a long time but there's been a lot of advances there I mean do the atmos is everywhere now and um we have customers that basically want to know how good a job they're doing at reproducing um The Source material the sound objects and you know can we come up with a way to say if I encode a source you know I don't know a violin a musical instrument at a particular distance a particular angle asmith and then I play it back on my quote surround sound system or my 7.2.1 whatever Channel system is it really doing an accurate job of reproducing the spatial audio um location and I I talk to these customers and they're I say so how are you measuring this and they're like we we're not we're we're just subjectively listening to it um some of them are use a laser pointer and closing their eyes so there's there's a big demand I feel to how can we come up with a metric to accurately determine if the sound reproduction of the spatial audio is indeed what the uh the musician or the recording artist or the mixer intended so um you know it's early days for us in that we were pretty familiar with uh interal cross correlation and interal level difference kind of measurements and they're they're very good at helping uh determine um a good job of localization as well as development but they're still not really there to say hey that's where this source is that's where the source is and and take into account the uh the room effects as well is really a big challenge so I'm hoping toh work on that and I had a good taste of this when I went to the as spatial audio conference in England which was sold out like immediately it's like the interest in this is huge not just from the people creating it but also from the people trying to reproduce it okay all right so that's quite a bit of uh information for today but uh maybe just a kind of a fun fact to kind of finish things off so what's maybe you want to share with us like what's the most interesting thing that you've had to test um I guess yeah sure sure I mean uh I I've been very fortunate that I've met a lot of different types of customers with a lot of different types of applications and people trying to measure some of the strangest things I mean it's the usual stuff right headphones smart devices phones watches Soundbars but you know some of the more interesting ones were um during the pandemic uh measuring um mask um I mean we all experienc the muffled sound of people talking through uh n95 mask and things like that and um I don't know if you guys ever saw the Dyson uh headphones yeah even know where that people started putting microphones outside the mask and then piping it into the headphones because frankly the uh it was very hard to hear speech intelligibility through a mask U one of my favorites though was uh we actually had a uh potato chip manufacturer who wanted to test or or company who wanted to test the crunch Factor potato chips and you know obviously potato chips are not don't crunch very well so speak of Crest factor and transients that we actually were able to use some of the same techniques that we use for capturing transient Distortion and uh I guess last but not least um I am a bit of a Gearhead and I I love cars and uh you know I I will admit I uh I own a Tesla and one of my complaints my first Tesla was you know it was super quiet when you're going around town but on the on the highway it was a lot of wind noise even though um the car is incredibly quote um you know low drag coefficient but there was a lot of wind noise so when um my lease was up and I got the new Tesla I was supposed to be quieter and I wanted to see if I could measure it I also wanted frankly to test out our new um audio interface that was portable and can be powered off of a USB um input so it didn't need to be plugged in so I went out and did some road tests before I got the the new Tesla and then afterwards and the new one supposedly had active noise reduction which also is kind of a big thing right now automotive industry and I wanted to see how good it was and frankly it wasn't great um so we'll be uh doing a u YouTube video on that hopefully soon I mean there were some other things that we measured which was pretty entertaining like uh how they uh how loud this Max SPL in the car and you know in the Tesla it goes up to 11 like in spinal tap and some of those funny things but uh yeah you're welcome to check check some of that out some of the those measurements are already and videos are already on our YouTube channel okay yeah we'll we'll put those links in our description for this one so that's stuff uh Simon do you have any questions before we I think I spent all my questions mate all right okay so that's that's great encourage everybody to like subscribe share and hit that Bell notification for future videos um Steve thanks so much for your time today this is extremely informative um and uh and well planned out we love the effort you put into the presentation today so thank you so much you're welcome thanks for having me and and then thanks for your patience no problem all right so see you next time bu okay by take care
2023-11-27 08:18
