Across Acoustics

A Different Way to Look at Soundscape Data

March 09, 2023 ASA Publications' Office
Across Acoustics
A Different Way to Look at Soundscape Data
Show Notes Transcript

Is there one "true" perception or assessment of a soundscape, or is it actually a combination of many different perspectives/assessments? How does a researcher represent data for a multifaceted view of soundscapes? In this episode, we interview Andrew Mitchell of University College London about his article, "How to analyze and represent quantitative soundscape data," (JASA Express Letters, March 2022), which addresses these questions.

Associated paper: Andrew Mitchell, Francesco Aletta, and Jian Kang. "How to analyse and represent quantitative soundscape data." JASA Express Letters 2, 037201 (2022); https://doi.org/10.1121/10.0009794

More on Soundscapy:

The Rest is Just Noise podcast: https://www.justnoisepod.com/

Read more from JASA Express Letters. 

Learn more about Acoustical Society of America Publications 

 Music: Min 2019 by minwbu from Pixabay. https://pixabay.com/?utm_source=link-attribution&utm_medium=referral&utm_campaign=music&utm_content=1022

Kat Setzer  00:06

Welcome to Across Acoustics, the official podcast of the Acoustical Society of America's Publications Office. On this podcast, we will highlight research from our four publications. I'm your host, Kat Setzer, Editorial Associate for the ASA. Today we'll be speaking to Andrew Mitchell of University College London about his article, "How to analyze and represent quantitative soundscape data," which appeared in the March 2022 issue of JASA Express Letters. Thanks for taking the time to speak with me today, Andrew. How are you?

 

Andrew Mitchell  00:39

I'm doing pretty good. Thanks very much for having me.

 

Kat Setzer  00:42

Yeah, no problem. So first, tell us a bit about yourself and your research background.

 

Andrew Mitchell  00:47

Sure. So I'm currently a research fellow at University College London, where I've also just finished my PhD. For my PhD, I was working on machine learning for soundscapes and trying to build a model to predict soundscape perception. One of my key focuses has been sort of trying to make a soundscape approach possible and useful for engineers and designers, trying to bring us just out of sort of research and looking at soundscapes into how do we actually make use of this for design? And I think like a lot of acousticians, I got into sound and all that sort of stuff by being interested in both physics and music. So I got my bachelor's in physics and music from Cardiff University, before then working as an acoustics consultant in Wales and in Los Angeles. And then I moved from there, from engineering and working in industry as a consultant, into research at UCL.

 

Kat Setzer  01:45

Awesome, very interesting. So what is a soundscape?

 

Andrew Mitchell  01:48

It's always a very good question. So there's a technical definition of it, which comes from an ISO international soundscape standard, which is a soundscape is the acoustic environment, as perceived by a person or persons in context. I think there's a little bit of question about whether this sort of accurately describes how people use the term. We can think of, like, underwater soundscapes, where there's no person there to perceive it, but we still use soundscape as the as the word from that. So for me, iit's about sort of a holistic consideration of how people perceive sound environments with a focus on considering both the pleasant and the, the annoying side, positive and negative sides of it, and trying to consider all the different things, all the different factors that might go into how you would perceive a sound environment.

 

Kat Setzer  02:44

Okay, that's very helpful. So what is ISO 12913? What does it say with regards to studying soundscapes?

 

Andrew Mitchell  02:51

Yeah, so as I mentioned, ISO 12913 is the current international standard on soundscapes. It's split into three different parts. So Part One sets out that definition for what a soundscape is, and what sorts of things are considered with it. Part Two sets out some recommendations for how you measure or assess the soundscape. And then Part Three gives guidance on how to analyze the data that you get from Part Two. And as I sort of said already,  the main point of this ISO standard is that soundscape studies focus on people's perception of a sound environment, take into account all the different factors, which is the sound itself, visual contexts, the people's own psychological factors, that sort of thing.

 

Kat Setzer  03:39

Okay, got it. So how do you analyze the perceptual attributes of a soundscape?

 

Andrew Mitchell  03:45

So, there's a really... There are quite a few different ways that you can analyze soundscapes, right? The main one that I use, and which is one of the different methods outlined in Part Two of the ISO, is called the "soundscape circumplex." One of the main things that all of these methods do for us is they give us a language in which to talk about soundscapes and sound perception. And particularly, they allow us to provide that language to non-experts that we're talking to, so that we can understand how they are perceiving the sound environment. So soundscape circumplex. Soundscape perception can be described along two primary dimensions. The first is pleasantness, and the second is eventfulness. And we can we can describe this in literally in two dimensions on a Cartesian plane, x-y axis, that sort of thing. So we put pleasant on the x axis, and we put eventful up on the y axis orthogonal to it. And we've got two ends of each of those. So we have positive x is positive, negative x is annoying. Positive y is eventful, negative y is uneventful. Okay, so we have neutral right there in the middle at the origin. So that's sort of the starting that is this two-dimensional plane of pleasant versus eventful.

 

Kat Setzer  05:11

What do you mean by eventful?

 

Andrew Mitchell  05:14

That's, that's all... That's actually one of the most common questions. 

 

Kat Setzer  05:17

Okay. 

 

Andrew Mitchell  05:19

So the best way I have to think about this is actually with what we call the secondary dimensions. So in additional, in addition to that x and y, we also have secondary axes at 45 degrees going in between them. And those secondary ones are vibrant, calm, monotonous, and chaotic. And one of the best ways, actually, for me to think about what eventfulness is-- because it is a little bit of a sort of abstract concept applied to soundscape-- is that actually vibrant, which sits right between, on that 45 degree between pleasant and eventful, vibrant... a soundscape is vibrant when it is both pleasant and eventful. 

 

Kat Setzer  06:04

Okay.

 

Andrew Mitchell  06:05

And a soundscape is calm when it is both pleasant and uneventful. 

 

Kat Setzer  06:09

Okay.

 

Andrew Mitchell  06:10

So eventfulness is kind of like, you know, this attribute of how much stuff is going on in the sound environment.

 

Kat Setzer  06:16

Got it. Okay. 

 

Andrew Mitchell  06:17

All right. So we have these, we have this dimension, and we have eight attributes that go around it: our primary ones, and then our secondary dimensions. That results in eight attributes, which are like pleasant, vibrant, eventful, calm, and so on. And the way that we then collect data  using that is to create a survey that we have people fill out. In my own research, and in quite a few soundscape studies, we do that in situ. So we go to whatever soundscape we want to analyze, whatever space we want to look at. And we either bring people there to fill out the survey, or we stop people as they're moving through and have them fill out this survey answering how they perceive that current sound environment on those eight dimensions.

 

Kat Setzer  07:07

Okay, got it. So what was the goal of this current study then?

 

Andrew Mitchell  07:10

So the main goal was to sort of examine the tools that the ISO gives us for how to analyze that data that we've just gotten from these questionnaires. What kind of came out of that then was actually a new analysis and visualization method that we've developed for soundscape data. 

 

Kat Setzer  07:33

Okay. 

 

Andrew Mitchell  07:33

And so, the main main goal this paper was both to sort of highlight a few of the issues that I've identified with Part Three of the ISO, but, perhaps more importantly, to sort of present this this new way of thinking and communicating about soundscapes.

 

Kat Setzer  07:50

So then one thing you talk about in your paper is how to really understand a soundscape you need to consider multiple people's perspectives of the soundscape, not just one person. So how do you do that? I guess it's probably the surveys. But can you go a little bit more into it?

 

Andrew Mitchell  08:05

Yeah. So a big part of my research has been sort of taking some of the work that had been done going into the ISO, which was, which tends to be very focused on understanding a single person's perception of a sound environment, their soundscape. And instead, particularly for this sort of engineering and design context, my focus is now on, I want to understand the soundscape of a place, right? We go to a park, what is its soundscape? Well, the reality is that the soundscape of a park is not just any single person's perception of it. It's made up of the perceptions of all the different people who use that park. 

 

Kat Setzer  08:45

Okay.

 

Andrew Mitchell  08:46

They might agree, they might disagree, one person may think it's very calm, another may feel it's neutral or maybe even a bit boring. And so the soundscape of the park, is therefore what we've now called the collective perception of all the people who use it. 

 

Kat Setzer  09:03

Okay. 

 

Andrew Mitchell  09:04

So that's, that's what we're trying to measure.

 

Kat Setzer  09:07

Okay, so you started talking about this a little bit--what are the limitations of the ISO that you see at this point?

 

Andrew Mitchell  09:13

So all of the sort of circumplex and that sort of stuff comes straight from the ISO. I think it provides some really useful basis  for those of us in soundscape studies to sort of consistently talk about what we're doing. And it has some some innovative methods. But the the issue that I was mentioning is that these were generally focused on understanding individual people's perceptions. And I'm much more interested in understanding the soundscapes of places. You know, if we design a park, how will people generally perceive it? Will it be described as a calm soundscape in the park or is it going to be monotonous? There are some other things that that come out of the ISO of like the methods that they give for sort of giving the central tendency, or the distribution of people's responses. the methods recommended there, to my mind don't, aren't really suited. And it also doesn't match how those of us in the field have generally been analyzing this data. So to really move forward and start thinking about things both of sort of higher-level ideas of the, what the perception of a space would be, how we can design for many people to use it, those sorts of things, I think we really need a method, which is more focused  on this collective perception, this broader looking at more people and looking at, to my mind, particularly the people who are actually using that space.

 

Kat Setzer  10:43

Yeah, that all makes sense. So how do you propose researchers discuss or represent the soundscape assessments that are collective?

 

Andrew Mitchell  10:50

Yeah. So the starting point, as I said, is we want to consider that variability. Right? We need a method that sort of inherently considers the natural variation, and allows us to discuss both the general agreement--that sort of central tendency, is a soundscape calm or vibrant or monotonous?-- but also the degree of agreement, or the sort of variation of responses, you know, is the soundscape really ambiguous? Do people just not agree at all? Or is the variation really small, and everybody has a really high degree of agreement? 

 

Kat Setzer  11:27

Okay. 

 

Andrew Mitchell  11:28

So the first thing that we do is we make use of this soundscape circumplex, right. And one of the ideas is you can take those eight perceptual attributes that we have measured in our questionnaires, and there's a method to sort of project those down to a single scatter point. Okay.  So, you can, you can essentially calculate a coordinate for each questionnaire, and plot that coordinate within the two-dimensional, pleasant-eventful space. So that means when we have, say, 100 responses within a space, that we can create a scatterplot of the soundscape. We can, we can start to show with that the distribution of everybody's responses, and the degree to which it was pleasant or annoying, or that, or all that sort of stuff. Then once you have that, scatterplot we can treat it like a density function. Right? So we can start talking about the distribution of those responses.

 

Kat Setzer  12:30

Right. 

 

Andrew Mitchell  12:31

So we can take that scatterplot calculate a bidimensional kernel density estimation, and essentially just put a density heat map, overtop of that scatter. 

 

Kat Setzer  12:42

Okay. 

 

Andrew Mitchell  12:42

And that really does a to my mind does a much better job of reflecting how that that range of perceptions that people have of that soundscape.

 

Kat Setzer  12:54

Okay, how can folks use the soundscape circumplex you talk about?

 

Andrew Mitchell  12:59

Yeah. So once we have that sort of heat map density, that's where it sort of gets really exciting to just to look at it, because now we can start talking about sort of the shape of a soundscape within that circumplex. We we have like sort of this little blob. And in some ways the full density heatmap gets a little overcomplicated if you have like a gradient for each sort of decile of distribution. So we simplify that down to plotting the 50th percentile. So if you plot just a single 50th percentile, you have this little blob that is within the circumplex. What that lets you do is now you have a better understanding of how that soundscape is distributed throughout the circumplex. But also you can start to compare. So we can compare the shape and the location of different soundscapes and see how much did they overlap? What percentage of this soundscape is in the calm quadrant? What percentage is in the monotonous quadrant? And that once you sort of wrap your head around that two-dimensional, pleasant-eventful space, once you get that lodged in, it then becomes a really sort of intuitive tool for understanding, communicating, and comparing soundscapes across different spaces.

 

Kat Setzer  14:21

Right, because you can just use your little graphical blobs and say like this graphical blob is skewed this way versus that way. And yeah, okay.

 

Andrew Mitchell  14:28

Yeah. And one of the difficulties for me has been, how do we communicate with like architects or designers who don't always think about sound all that much, right? And, sure, we can we can talk about DBA values, we can we can talk about CDL or LDN values and all these different things that we've created to summarize environmental sound, but it's not often that sort of intuitive for a lot of designers.  And soundscape has already tried to do some of that by now talking about it in language of perception of pleasant or vibrant, that tends to work a bit better, but then it always becomes a bit fuzzy. And it's hard sometimes for people to latch on to what all that means. And particularly to compare. How do you how do you start comparing things in that case? And that's what I think the this visualization in particular, can really help do. It has lots of... the general concepts have lots of, lots of other uses, you know, just having that scatterplot of coordinates, makes your analysis in all other cases,  whether you're using different metrics, or even if you're just doing like t tests or something like that, that becomes quite a bit simpler, using those coordinates, as opposed to using the eight attributes. But the visualization itself really does get the message across with like, sort of this nuanced description of the soundscape of the space. Right. 

 

Kat Setzer  15:59

Yeah, that totally makes sense. So then in your article, you also mentioned that in some cases, simpler metrics are needed for analyzing soundscapes. What do you suggest?

 

Andrew Mitchell  16:08

So certainly, it's still difficult to just work with these densities. Some of my work is has moved towards doing that, taking sort of a Bayesian approach to these density distributions. But we certainly still need simpler metrics to summarize them. So the nice thing about this is, these are now we've taken what are previously Likert scale responses, so just 1, 2, 3, 4, or 5, for each of those eight attributes, which can be sometimes a bit more difficult or unfamiliar to do statistics on. And by taking the scatter, these scatter plots, or scatter coordinates, we now have continuous variables. Right? 

 

Kat Setzer  16:53

Okay. 

 

Andrew Mitchell  16:54

They're, they're kind of just quite familiar to work with. So we can use all of our usual summary statistics of mean, standard deviation, skewness, that sort of stuff. We can also take a page out of the Community Noise Annoyance book, and discuss things like the percent likely annoyed, or likely percentage of people who find the soundscape vibrant, or that sort of thing. Our team is now also working on developing a single index metric to summarize the quality or the success of a soundscape, which is based on this visualization method. So, hopefully, in the future, we'll have that sort of thing as well.

 

Kat Setzer  17:30

Got it. So what are the limitations with your proposed methods?

 

Andrew Mitchell  17:34

So one is kind of just about changing mindsets, right? So both within the soundscape community and in the machine learning and mathematics side, the sort of thing that I've been describing requires a bit of a change in how you think about this variation. So often, people will think that there's like a ground truth assessment for a soundscape. And the variation around that is measurement error. Right? So you get this spread of responses, and somewhere in there is "the truth," and then you have error around it. Sure, there certainly is some measurement error, like, absolutely no denial of that.

 

Kat Setzer  18:12

Right. 

 

Andrew Mitchell  18:12

But really, to me, the variation in people's responses is the ground truth. It's, it reflects the diversity in people's perceptions of the spaces, and it shouldn't be thought of as error. It should be thought of as, that is the soundscape itself. That spread is the ground truth. 

 

Kat Setzer  18:31

Yeah. 

 

Andrew Mitchell  18:31

And that becomes difficult both in how you think about it soundscape-wise, it also, it can make some of those statistics more difficult, you know, maybe the mean, isn't the right thing. Because just getting a single point for a soundscape doesn't, it's not reflecting the reality of how people actually perceived it.

 

Kat Setzer  18:50

Right, because it can be paradoxically vibrant for one person and chaotic for another.

 

Andrew Mitchell  18:58

It also it goes to how we have to treat outliers as well. You know, I tend to take a sort of general approach in survey work that very rarely should we be excluding outliers in surveys. You know, if we have some evidence that the person clearly didn't understand the questions, which we can do that in soundscape surveys... so if someone answered, I completely agree that the soundscape is pleasant. I also completely agree that it's annoying, I completely agree that it's vibrant. I completely agree that it's calm. They probably didn't understand the questions. And so if they answer the same all the way down, we can kind of, we can dismiss that.

 

Kat Setzer  19:39

 Right. 

 

Andrew Mitchell  19:40

But if they've just answered completely differently from someone else, that's, that's an outlier, but we shouldn't exclude it. That's their perception, it's their truth. And maybe there's enough people who would perceive it that in that outlier sense, that now becomes a, you know, a significant proportion of people. And we should consider that and notice it and go, "Oh, what is it that's making those people perceive it very differently? Was it a different period in time?" 

 

Kat Setzer  20:13

Right.

 

Andrew Mitchell  20:14

That's very common. One of the examples I gave in the paper, is, we were doing a survey in Regent's Park here in London. Very nice, very big park. And we're getting the responses that we expect, generally calm, a little bit vibrant, overall pleasant. And then if you just look at our data, you see this little blob of people who say that it's very chaotic. And if you're just looking at that, in sort of a statistics way and said, "Alright, I'm going to exclude some outliers." And those will, those will get the outliers. Or if you didn't exclude it, it would shift your mean, and it would change your response. Well, the reality is that we had helicopter flyovers at that period. So because we were out there for several hours doing these surveys, and actually over several days, they actually had a different soundscape. They had a different sound environment. In this case, it was actually it was rehearsals for I think, for President Trump's visit to London. So that was fun. Yeah, if we had just sort of taken as sort of, alright, we exclude these that clearly isn't right, we would have missed that, or if we just included them, that also would have been incorrect. But when you plot it on the scatterplot, that becomes really obvious, right? And really helps you see what's actually going on over time, and through different people in that soundscape.

 

Kat Setzer  21:37

Right, and also to... it makes you more familiar with a soundscape's... what could happen within a soundscape, right, since they presumably won't be static.

 

Andrew Mitchell  21:46

That's another thing just on a visualization side. Depending on how exactly you want to represent the... so the examples that are given in the paper of these visualizations, usually, these are surveys that were taking over several hours and several days, so they have variation, which isn't due to people perceiving it differently, the same thing differently. It's just the variation of that soundscape over time. So we thought of that as part of the density. Another option would be, you can have like a little animated thing or track in the circumplex, showing the soundscape was here and then it moved, and then it moved, and then it moved, and then it moved. And now, that's a much better way to show the change of the soundscape over time.

 

Kat Setzer  22:28

Right. Okay, very cool. Do you have any other thoughts you'd like to add? 

 

Andrew Mitchell  22:33

Sure. So the first is if people would like to try this method out, or to look at some of the data that I've been talking about, I've developed an open-source Python package, it's called Soundscapy, which is specifically for performing this sort of analysis. We've also released a large database of soundscape surveys for people to try it out on, which is called the International Soundscape Database, and it's freely available open source. So I highly encourage people to go to the Soundscapy GitHub repo and take a look at some of the examples and tutorials that I put up there.

 

Kat Setzer  23:04

Yeah, and we will link to those in our show notes.

 

Andrew Mitchell  23:07

Yeah. And then, if as people are doing that, and if they read the paper, I'd love to hear feedback, particularly from working consultants and architects, for how useful and understandable they find these plots, because, like I said, a key goal was to help us improve communication between acoustics and soundscape experts and the designers, architects, and policymakers who sort of actually have direct input on how these spaces are designed.

 

Kat Setzer  23:34

Yeah, of course, totally. Well, thank you again for chatting with me. You make some really interesting points about how researchers think about and discuss soundscape data. So as I just mentioned to folks, for our listeners, if you're interested in learning about more about Soundscapy, Andrew's Python library for analyzing and visualizing soundscape assessments, we'll be linking to it in the show notes. And if you're interested in hearing more about the science of soundscapes, you can also listen to his podcast, The Rest Is Just Noise. Have a great day.  

 

Thank you for tuning into Across Acoustics. If you would like to hear more interviews from our authors about their research, please subscribe and find us on your preferred podcast platform.