Episode

Sure you had to be a good Bayesian.

Alex Andorra: Hello my dear Bayesians.

Today we're talking about voices, vowels, and everything your brain infers every time someone opens their mouth.

My guest is Santiago Barreda, professor of linguistics at UC Davis, phonetician, and co-author with Noah Silbert of Bayesian Multilevel Models for Repeated Measures Data.

A book written to get linguists doing Bayesian stats on the data they actually collect.

We start with how Santiago came

to Bayes, and why he thinks students should jump straight to multi-level models instead of spending a year t-tests.

We also talk about math anxiety and why learning statistics is more like learning the piano than memorizing facts.

Then we dive into his research, why listeners are often wrong about who's speaking, but wrong in predictable ways, and how vowel normalization works a lot, like seeing that a car

is far away rather than tiny.

Along the way, we cover red deer, koalas, how languages change, and David Reby's idea that speech may have evolved from catching cheaters.

A big thank you to Kristin Predeck for connecting us.

This is Learning Bayesian Statistics, episode 168, recorded September thirtieth, 2026.

Show you how to be a good Bayesian and change your predictions after taking information.

And if you're thinking I'll be less than amazing, let's adjust those expectations.

What's a Bayesian?

It's someone who cares about.

Alex Andorra: Welcome to Learning Bayesian Statistics, a podcast about Bayesian inference, the methods, the projects, and the people who make it possible.

I'm your host, Alex Andorra.

You can follow me on Twitter at Alex underscore andorra, like the country, for any info about the show.

learnbayesstats.com is Laplace to be.

Show notes, becoming a corporate sponsor, unlocking Bayesian merch, supporting the show on Patreon, everything is in there.

That's learnbayesstats.com.

If you're interested in one-on-one mentorship, online courses or statistical consulting, feel free to reach out and book a call at topmate.io slash alex underscore andorra.

See you around, folks, and best Bayesian wishes to you all.

Santiago Barreda, welcome to Learning Bayesian Statistics.

Yeah, uh a pleasure.

Thank you so much to uh Kristin Predeck, who is a very avid listener, or at least I hope so.

Um She put us in contact because you wrote a very interesting and good book and influential book for her and the rest of your field.

Uh and we'll get to that book in a minute.

But before, as usual, you know, uh I wanna dive into your origin story, Santiago.

Can you tell us what you're doing nowadays and also how how did you end up doing that?

Santiago Barreda: a professor of linguistics at the Department of Linguistics at UC Davis.

Um I mostly focus on speech perception and phonetics.

So I'm a phonetician, you could say.

Which is basically the study of of speech, sounds, acoustics, the classification of those.

Um I got into linguistics initially as an undergrad.

I took a Latin class and I was just fascinated.

It just seemed unreal that this was a real language, that people just opened their mouths and and sounds came up because

In Latin you can have a free word order and you mark whether something's ver uh subject, object, or and so on, based on how the words end.

And so you could have this you have this really complicated grammar and it just seemed crazy to me that you could listen to a language and the words could come in any order and

you could make sense of it.

And so I got interested in how is this possible?

Languages are way crazier than I than it seemed to me, just from um Spanish and English, which are the languages I speak.

So and then early on I got interested in cross language vowel perception.

So when you hear a vowel, how do you know what vowel you're hearing?

And also when you hear a vowel in another language, how do you classify those vowels?

And so this is crucial to understand for understanding accents, especially second language accents, which are you can't produce uh the sound accurately, but also a part of this is

you have a really hard time often separating sounds in this new language.

And so I became really interested in vowel perception specifically.

This led me to the University of Alberta to work with Terry Nearey, who was a specialist in vowel perception.

Um I was very lucky to work with him.

He's passed away a few years ago.

But I really see what I'm doing as a a continuation of his work, which is a focus on normalization.

sometimes called speaker or talker normalization.

So in the most general sense, normalization is mapping.

From acoustics, so a physical sound out in the world, to some sort of linguistic representation or information.

So, you know, information being passed f uh f using the linguistic system to convey some sort of signal or message.

So here's the key problem that you could sum it up like this.

You and I can both say, ah or we could say hi and you could take a hundred people in a room and you they say hi.

And you record the acoustics, and the acoustics can be radically different.

And yet you play the sound for people and they'll say they're saying the same thing.

So the question is, physically very different different things are perceived by humans as being the same on some level.

Which means that sometimes acoustic differences don't matter for linguistic information.

But the thing is, those same acoustic differences are how we make different speech sounds.

So I say ha and I say eh.

And those are small acoustic differences that result in different sounds.

So sometimes differences cause linguistic acoustic differences cause linguistic difference.

Sometimes they don't.

And understanding which ones do and which ones don't is basically s referred to in linguistics as this the normalization problem.

And so I became interested in this problem, but really the the the most direct way to test this empirically is using

human listening experiments because fundamentally this isn't like can we make a computational model that guesses really well.

We we r are interested in well what are humans doing?

Whether they do it well or not, whether some AI can beat them.

We want to know, well if humans are making mistakes, why?

What kind of mistakes?

So to understand this you need perceptual experiments which we need you under you need to understand acoustics to some extent and signal processing to make your stimuli.

But then of course you get a bunch of data, a lot of decisions

by listeners, categorical or continuous in some s in some cases.

And then you need to you need to model them.

And that is often leads to relatively complicated models.

It's complicated from the perspective of the things linguists usually do.

And so this is where I'm at right now basically, which is manipulating acoustic stimuli, presenting them to people, collecting judgments, and

using that data to make inferences about how humans map acoustics to linguistic information.

Alex Andorra: Mm-hmm.

Okay, yeah.

Yeah, that is super fun.

Uh and is that something you like when does when did that interest, you know, come up?

Uh was that something that you knew you wanted to do pretty fast in your studies or was that something that ended up pretty randomly being the main interest of yours?

Santiago Barreda: Um during my masters I did the uh my main experiment was predicting classification of English vowels by first language Spanish speakers who had learned English as adults, like

in the late thirties.

and the problem I faced so the thing about vowel perception is many people in linguistics I think think that this is a solved problem or a simple problem.

But we don't

understand even the basics, to be honest, when you when you dig into it.

And I faced a very simple problem or what seems like a simple problem.

Um I have a bunch of different people's speech, male and female, that have the same dialect, or do they?

So there's acoustic differences.

How can you establish whether they're linguistically what I'm gonna call this phonetically different.

Phonetically I mean like the speech, the continuous speech information like a like the information in vowel sounds.

I don't really care if they're acoustically different.

It's like trivial that they're acoustically different.

I want to know if they're phonetically different.

For that I need to know when people are faced with two humans' uh vowels, do they think they're different or not?

And for that you need to know what people hear sort of h inside their heads.

What and not the acoustics.

And so I was faced with this very simple problem.

How do I compare these different people's vowels at the at a very early stage of my research and

When I started looking for answers, I realized there really weren't any about this very s simple problem and I would call it foundational, is which is s you know, make highlights

the the lack of research into it.

Almost everybody that collects speech data needs to compare different speakers.

but we don't currently know exactly how.

And this led me to University of Alberta because uh Terry Nearey was one of the I mean probably the leading expert in that topic.

Alex Andorra: Okay.

Yeah.

Yeah, that makes sense.

And is that at that time that you started getting into Bayesian stats or actually what what drew you to Bayesian stats specifically?

Santiago Barreda: Initially I I was doing all frequentist stuff.

but i I'm also talking about like twenty years ago at this point, which is kinda crazy.

But back then, you know, two thousand six, two thousand eight, Bayesian I know people were doing Bayesian stuff, but it wasn't nearly as as accessible as it is now.

Like n nothing of the sort.

Uh and it really wasn't used.

It really it's kind of not used now.

It's it really wasn't used back then.

So I was doing traditional stuff, ANOVAs, T tests, things like that.

Then people started using

um LME lmer, the Elmer function, or some people call it Lemer, I think, in the lme4 pa uh R package.

So I started using that.

And then the main problem was that it doesn't converge.

This is a classic thing.

Convergence issues, convergence issues.

And then it also just really doesn't work for complex uh multi level, like repeated measures type models.

So like higher what are called random effects models, mixed effects models, hierarchical hierarchical

models, however however people want to refer to them.

Um I just it just couldn't converge.

It w it didn't give me the information that I wanted and and needed.

Uh and so I started doing Bayesian stuff in two thousand and fifteen or so.

my friend Noah Silbert got me into it.

He's the person I co wrote the book with.

Uh as a purely practical matter.

I just see I'm a I'm a pragmatist.

I just needed models that worked.

And it wasn't like philosophical

um sort of zeal that uh got me to use Bayesianism, but but as I started to use it, I realized that I did agree basically with the philosophical positions of Bayesianism.

Uh I liked all the information that it gives you.

I like the flexibility uh that you get that you can have by having different kinds of error distributions, different kinds of model structures easily within the same framework.

Um and it also lets you weird models that don't exist.

What I compare it to is when you're a frequentist, and I know this is a simpli gross simplification, but it applies I think to most sort of normal people that aren't gonna be

right coding their own like uh specific d special distributions in Python or or stuff like that.

Um when you're a frequentist, it's like going to shop at IKEA.

So like you Ikea has a bunch of great stuff, but

You know, if they don't have the bed that you want, well, what do you do?

Well you buy them the closest thing to the bed that you want it.

Because that's what IKEA has.

And when you're a Bay when you're doing Bayesian modeling, it's like learning to be a carpenter.

And you learn how to make legs using a lathe.

And then you learn to make corners.

I don't know.

I don't know about carpentry, but the different kinds of corners and stuff, you know.

The elements of carpentry.

And you put them together and you learn to make exactly the model

And but I I don't want this to seem like I'm saying like let's overfit our stuff.

I'm saying a model that reflects what you think is really going on in your data.

That's a a better representation of our research questions and the things that we're specifically interested in, instead of going the other way and sort of ramming our data

into a pre fit standardized model structure.

Alex Andorra: Yeah.

Yeah.

Yeah.

I think it's a great great description, great comparison.

I remember a few years ago I used to use the Playmobil versus Lego analogy.

So yeah, it's like yeah, frequentists would be Playmobil here.

It's it's great if you have if you happen to want to play with exactly that kind of um situations, but if you want something custom, Lego is Lego, Lego, let's say Lego.

Lego is gonna be better.

So but I think your your analogy is even better.

So so yeah, I'll I love it.

And I'll I'll definitely shamelessly steal it.

and so you actually brought up your book with Noah Silbert.

and so the book is called Bayesian Multilevel Models for Repeated Measures Data.

I put the link in the show notes.

And um so yeah, I think it's great to have you here to talk about it because it's

It's a very it's the kind of book that I think is very important and helpful because it tries to do that translation of okay, why would you even care about Bayesian statistics but

not from a philosophical standpoint, just from a practical standpoint, if you are doing linguistics research or just in general repeated measures, data, and well, you have to

model things, you know.

So

Yeah, can you tell us more about why you wanted to write this book and for whom mainly?

Santiago Barreda: Right.

So in in linguistics and similarish fields, you know, psychology, speech sciences, things like that.

Basically all data is repeated measures data.

Certainly in linguistics, you know, you don't go to the trouble of collecting speeches with somebody and have them say one syllable and then send them home.

And then you don't also put them in a booth and say, Hey, what does this word sound like to you?

They say one thing and then they go home.

So usually you have dozens, but actually closer to h hundreds of trials per listener, and then you have a a a pool of speakers.

So

We always uh spend time sorry, we always have repeated measures data.

And we shouldn't be spending so much time on non repeated measures models, basically.

Because the way we usually train linguists so I I'll only speak to linguistics because this is the only thing I know anything about basically.

I can't really speak to how things are done in other fields, but it seems to me that this is very common in linguistics and it wouldn't surprise me that this if this were common in

other related type fields.

When we teach students statistics

we have them go through like a a long sequence of traditional the traditional sequence so they'll look like one sample t test, paired t test, ANOVA, two way ANOVA, and then they

build up.

But most of the things they learn are first of all, in my opinion, more confusing than the like the analogous or equivalent Bayesian approach.

And they can't use so we're asking to spend like a year plus on

Concepts that they will think sorry, models that they will never be able to use.

So my thinking was can't we just jump straight to the stuff that they will be able to use?

Which, for example, like instead of just a one sample t test, I teach in the first thing we cover in the book is uh the the multi-level equivalent, which is like a one group of

list.

So we have one sample t-test, but within each listener we have a bunch of data.

So it's kind of one sample at the second level.

It's not I don't think it's honestly even more complicated than a T test because we don't have to talk about the reference distribution and the the test statistic and all this

other stuff, degrees of freedom, all this other stuff that is not super intuitive to people.

And so we c I thought first of all, we could just jump to the interesting thing.

So no what I say in the book is no more boring boring origin story.

Spider-Man doesn't have to hit get bit by the spider at the beginning of every Spider-Man movie.

Like everyone knows how it starts.

Let's just get to the good stuff.

So I try to get to the good stuff in the book.

The other thing is, this is a little outdated at this point, but 'cause the world has dramatically changed even in the last five years.

When I started working on this book, there was no generative AI.

And I was working with JAGS.

Sorry, when I before uh before uh brms came out.

Not when by the time I started with this book I was working with brms.

But

was just really difficult for me to figure out how to write the very complicated JAGS models I had with basically no examples of these.

There's a lot of really great books, especially at the time, like you know, Gelman and Hill's uh black and gold book, Kruschke's doing data analysis, obviously Statistical

Rethinking, do but these books really are are very general and also devote a s very small part to repeated measures type designs.

And mostly focus on very simple things like bivariate regression repeating measures type designs and don't sort of expand on them to the sort of complexity that even very simple

um experiments in linguistics require.

So what I wanted to do was give a lot of examples that people can copy fully worked from what research question can we address with this model?

How do we structure the model?

How to interpret the model?

What plots to make?

How to talk about the results.

And then at the end of the earlier chapters a comparison to the analogous frequentist analysis, just to sort of reassure people that we get the same results, basically.

With analogous models, we get largely the same results for the amount of data that we have.

There's i what I usually tell people is if you get dramatically different results, it's usually a sign something's gone wrong.

with the with the sorts of data sets that we have, in phonetics.

Alex Andorra: Yeah.

Yeah yeah.

Yeah.

And what's what's been your your experience so far?

Uh you know, uh where people how have people received the not only the book but this new way of thinking for them?

Have have they been able to adopt it or do you see a lot of resistance to change?

And if yes, how how do you handle that?

Santiago Barreda: be honest I've never once I've been working more or less exclusively with Bayesian models unless you know like I said I'm a pragmatist so if it's a very simple thing and it can be

done quickly and cheaply with uh bootstrapping or frequentist approaches I'll do that.

But um I've been using it exclusively for about ten years and sorry, n almost exclusively for about ten years and I've never once had a negative comment from a reviewer or anything

like that.

So I maybe I've been lucky.

But

Um I've had a very positive experience and I've also had a really positive experience with the class.

So I teach a class here at UC Davis, a graduate class for linguistics students using the book.

And to be honest, I kinda wrote the book because I wanted something like it for class.

Because and I there's this quote I really like that we treat others as we would like to be treated.

Sort of for better or for worse, we do that.

And I kind of think that this is I

I wrote the book that I wish I'd had basically when I was learning.

So it's kind of written for the way I teach or learn.

with and and so it fits naturally with the the class I teach.

Uh and I used to teach this test 'cause I've been teaching it for for ten years here, with a frequentist two-way ANOVA, stuff like that.

And I really get w much better results.

Students pick up way more and are able to do

way more in just the ten weeks of the classes, because unfortunately it's only ten weeks long.

Uh jumping straight to Bayesian multi level models than spending a bunch of time l like I was saying earlier, introducing things that they won't actually use, uh that are sort of

uh not as intuitive is what I would say.

Alex Andorra: Yeah.

Yeah.

I mean Yeah, that's been my experience too.

Um honestly both when when teaching and and in the industry.

Um something I also I've also seen all the time and I still see all the time is students or just people in general getting um discouraged or even not trying

these kind of models and methods because they convince themselves that they're not good at math, you know, before they even try to to do the model, to actually write the model um

that they could use in their own research or their own project.

And I know it's something that that is also dear to your heart.

So how did how did that shape the way the book is structured?

Santiago Barreda: I think this uh I notion that people have that they're not good at math, I really think it's because so obviously this is just my opinion, but I think people have a mistaken

understanding of or at least a not constructive understanding, let's say, of what it means to be good at math.

Because you know, there's obviously like levels to it.

So most people won't be Euler, Gauss, Hilbert.

Mm-hmm.

But that's like saying, you I'm bad at piano because I'm not Mozart.

That's a little bit of a silly thing to say.

So if we hold ourselves to a lower standard

I think that people think that math is like about raw brain power intelligence and then if they're presented with a mathematical concept and they don't immediately understand it, it

must be they lack that intelligence and then won't be able to learn it.

Alex Andorra: Yeah.

Santiago Barreda: W the way I pitch it in class and in the book is I think statistics math is much, much more like uh playing an instrument and learning a foreign language than it is about uh

declarative knowledge or rote memorization.

So nobody would

Pull out a new piano piece, try once, fail, and then think, I'm bad at math.

Yeah, exactly.

Somebody would be presented with irregular verbs or you know, subjunctive or things like this and think I will never learn Spanish.

So this is silly b behavior, but people I really do and I've I've asked my class to please not do this, will read a chapter of a math textbook once, will not understand it, and will

think it's because they're bad at math.

So what I would like to encourage people to do is to not take this perspective, to assume I tell students you assume you won't understand it the first time, and even maybe even the

second time, and then you'll come back to it a little bit later.

And then it will just make sense a little more.

And eventually and eventually things will make sense and it it won't be clear how or when, but a lot of the times I think the key is practice immersion.

We talk about immersion of a foreign language, but in in the math, there's this great line that's attra it's who knows if it's true, but it's attributed to John von Neumann where a

student says in uh about some concept a professor I don't understand.

And von Neumann says, uh young man, in mathematics you don't understand things, you get used to them.

And so I love that line because I think that fits in with this sort of practice and uh and procedural knowledge more than declarative knowledge.

And I think if von Neumann saw it like that, then you know we should definitely give ourselves a break.

Alex Andorra: Yeah.

Yeah yeah.

Couldn't agree more.

and yeah, that's always some I I always repeat also answer uh a variation on that, yeah, yeah.

Wh which is like well, if you start playing football and you're not um as good as Messi, uh does that mean you cannot play football?

I don't think so.

And uh and also the good thing is that, you know, contrary to professional sports or

um very elite kind of works like that, you actually don't have to be in the top point one percent of the population, which even more than that is not only very good, but it's also

just genetic talent, you know.

It's like genetic draw, extremely good naturally, plus a lot of work.

So yes, um the ceiling is super high, but that's the top

point one percent, you know.

The good thing is that in research or data science industry, things like that, you don't have to be at the right tail of the distribution to actually make contributions and do

great things.

Um most of the people will be by definition in the bulk of the distribution and they can make perfectly fine work, very good work, and also an extremely decent living.

um doing that.

Whereas yeah, if you try to be a professional sports athlete athlete, if you're not in the top point one percent, you're not gonna make a decent living and and that's gonna be a

problem.

Like there is no middle class there basically.

Um so yeah, like completely agree with you.

Um I I really love the the piano the piano image because you know this is it's really silly.

It's like you know, you try once and you're like, damn I'm s I'm bad.

I guess I cannot play the piano.

Santiago Barreda: It's silly when we talk about piano, but I really do think that students read a chapter of a textbook once and don't understand it and and make the same inference, which I think is

just as silly.

And so I put this all in the in the uh foreword of the book or the introduction.

I don't remember, but the the book is basically built around this idea because every chapter is structured like an academic paper basically.

It has like the

classic five part type structure.

Introduce the sort the problem, talk about how we can use the data and model structure to answer this, uh, discuss results, including interpretation and plotting.

and then do well, I think I skipped more, but conclude conclusion and sort of wrapping it up.

And so and another key thing actually that is uh that we use the same data set.

I want to mention this because I think this is very important.

I think it's really unusual.

I've never seen a book that does this, but uh I was just inspired to sort of s go back to the drawing board with the way stats books are approached because I I just wanted to do

something that I thought worked for the book.

And something that I find a little bit frustrating is that books often use a little toy example for every chapter or even just each concept in your work you're just dashing

through examples.

But I wanted to build up to these very complicated models.

So what we do in the book is analyze

this one relatively complex data set that supports all kinds of analyses because you could ask all kinds of questions with this data set.

Uh and call and work through the same data in iteratively more complicated analyses.

So you get gradual it's kind of like resistance training.

You know, you you go and you lift 20 pounds, then 25, then 30.

And then so we're sort of taking the same approach towards gr I call it unpeeling an onion, basically.

Like putting the idea and back back together and adding layers of complexity as we work through the chapters.

Alex Andorra: Yeah.

Yeah.

Yeah, yeah, I really love that.

and and I think it's a great it's a great way to to introduce people.

That also helps people like maybe be less intimidated because they maybe don't feel that they have to read the whole book cover to cover to actually get something from it.

Um so I think it's it's definitely a plus.

something I I discovered actually while preparing for your episode is that a lot of your research

is about how listeners judge a speaker's size, age or gender just from the th sound of their voice, which is yeah, I mean now that I think about it, I definitely do that

automatically, and that sounds like a very important inference to be able to make.

Um just even just from a I guess uh you know survival instinct perspective.

But yeah, can you tell us more about that?

I found that fascinating.

Santiago Barreda: Well it's interesting that you mentioned uh survival instinct because I think that's actually secretly at the root of all this, but I'll I'll leave that for uh for some l

later uh comments.

But the idea is you're right, we all make these inferences from voice.

Like someone could be listening to us speak right now and not see us and certainly know a bunch of stuff about this.

For example, we sound like adult males probably.

Um

You know, in linguistics accent just means pattern of sound, you could say, for example.

So everyone has an accent.

So they would hear both of our accents and they would make inferences about where we're from, where we're from, what our first language is, what are where we've lived, things

like that.

So I always wonder what I sound like to people.

I I'm from Peru originally.

I grew up near Toronto and now I live in California.

So I always ask students, where do you think I'm from?

And it's very interesting the things they say.

Uh and so we all make these these inferences.

But the the key thing is we use the same acoustic information to uh make inferences about what the person is saying, obviously, ling speech, linguistic information.

But we use the same information to make inferences about who the person is.

And these things are related, is what I would say.

That is the the most the center of my research is that these things are often treated as two things.

um that are independent.

One some people model one, some people model the other, but w you can actually see these decisions trading off.

So they're actually um not independent at all.

Alex Andorra: Hmm.

Yeah, yeah, yeah.

I mean that makes sense and that's I find that super interesting.

and actually so just from an anecdotal perspective, um like y I just realized I was in the back of my mind, I was looking for it since we connected on that chat, and your accent rem

the way you speak English

Your accent in English reminds me exactly of So have you seen um Friends, the TV series?

Okay.

So in in a few episodes, Phoebe is going out with I think his name is Gary.

He's a police officer.

And she ends up uh breaking up with him because he shots a bird.

And so you have the exact same accent as Gary.

This is like this is amazing.

I don't know where the accent is.

But I love it.

And so yeah, I'm very happy because I finally put put a face on the on the on your accent.

So that was just the anecdote.

But to get back to the serious part of the episode, um Yeah, like the that inference does does make a lot of sense, you know.

Like why why would you even do that?

Uh is that also related it's i like are these processes that happen in

in the brain in the way you would infer all these past and and just like characteristics of a person just by their voice.

Um is this related in in some way to you know this ability that we have to infer the words even like if we only see the the consonants?

You know, like I I love these the these experiments that they've made where you remove the vowels

And people can still read the sentences because your brain just infers the vowel.

Is is that related i or is that actually different or maybe that's not what you do and and you don't know?

Santiago Barreda: Well, I mean I don't really see them as related specifically, but I do I mean, I think the the bigger picture is that here's what I would say.

I see them as related in that w it's the same sort of guesswork from limited information using previous knowledge.

We're doing that all the time in speech.

It's w a lot more cheating, shortcuts.

guesswork stereotypes than than we might think.

That's another thing that I constantly realize in my work.

We th we kind of there's this traditional view of things as sort of optimal, smart.

And I I mean you c I think you could still say people are being optimal down their beliefs.

It's just their beliefs are weird and wrong often.

And so the cues are often used in sort of strange ways.

People are often wrong.

Because they're guessing, people are often wrong.

The interesting thing is that they're often precise.

And by that I mean like when listeners misidentify speaker information, they would usually be doing it in predictable ways consistently.

So other speakers would make the same mistake.

So it's for example, it's not that people can guess, for example, my ethnicity or race or sexual orientation accurately.

But if someone classifies me in a certain way, very likely other people make the similar types of mistakes.

So we consistently misidentify uh characteristics, basically.

Alex Andorra: So you mean when I like if I'm was the first one so so it's anchoring effect, basically.

If if my if my guess if the guess that I make about you is open, then other people will tend to have the same guess?

Santiago Barreda: Um sorry, I'm not that the other people would have heard you.

What I mean here's what here's here's this classic question.

Can people guess sexual orientation from voice?

This is the sort of thing I'm interested in because uh it's indexical information from voice.

Obviously this has to be social, right?

There's no I don't think any reasonable people are saying there's some anatomical difference according to sexual orientation between the larynx and the and the lips.

That would be very strange, unusual, I think.

So

Uh, can people guess?

The as all the research I've seen suggests, no, they can't.

This is like coin flip.

However, when people are misidentified, when a when a homosexual male is misidentified as straight, or when a straight male is misidentified as a homosexual male, most people many

people will make the same mistakes for that voice.

And so it's not that people can guess accurately, it's that they are predictable.

I can tell which voice is likely to be misgendered, for example.

I can tell which voice sounds small.

I can tell which voice sounds large.

People are are reasonably predictable from the acoustics.

Reasonably, of course, people are do all kinds of crazy stuff.

But people's right and wrong answers can be reasonably predicted from the acoustics.

Alex Andorra: Mm-hmm.

Okay.

this is fascinating.

Okay, okay.

Santiago Barreda: sorry.

Like, you know, then is a stereotype for example.

There's a classic, this is a classic example from sociolinguistics stereotypes.

People say New Yorkers are rude.

And so knowing that somebody's from New York, you might think might affect people's rudeness ratings of that person.

And so what we're interested in is on the level of not are New Yorkers actually rude, but

Do people think New Yorkers are rude?

And are people consistent in their guess that a New Yorker will be rude?

And so there's this there's a certain disconnection between the empirical reality of are they rude?

and do what do people think?

And it turns out people are super predictable in their stereotypes.

Alex Andorra: Yeah, yeah, yeah.

Okay.

Yeah, I mean, we're from the same species, so that that does make sense.

It'd be interesting if there were still another human species to see if they had different heuristics, you know, uh like that.

Honestly that'd be Well that'd be great.

Santiago Barreda: Be honest, we don't even have to go as as far as species because this is very strongly uh contingent on sort of social knowledge and information.

And so one interesting one interesting way to test this is to see how judgments from acoustics are made different based on different people's different beliefs about the

speakers, for example.

And so that's one way that that we can test the different ways that these

uh well I would call them beliefs I guess about the characteristics of people might vary across societies.

Alex Andorra: Mm-hmm.

Yeah.

Yeah.

Do you know just a random question like that that pops into my head, but do you know how and if that work is applied?

Because I can definitely see an application in like intelligence work where I feel like if you have if you have the ability to identify somebody at least somewhat accurately just

from their voice, if you're doing some intelligence work that that

that's actually can be that actually can be very helpful.

But yeah, I guess by definition we don't really know if that's applied.

But yeah, do you have do you have any other like do you have actual examples of of that work being applied somewhere?

Santiago Barreda: Well I don't know about applying it for well okay, first of all, identifying v uh speakers from voice is something I'm super skeptical of.

uh Actual speakers because you know, we know skilled impressionists do a really good job, for example, at a superficial level and there's so much acoustic variation within speaker

that I it depends what it would be used for, but if we're talking about like the certainty required to put somebody in jail or worse, then I I don't think that

I just I don't think it's a technological issue.

I think it's like an epistemological issue and and we should be really wary of of that sort of thing.

Of yeah of identifying individuals from voice.

I only say this because I know that sometimes it's used in criminal proceedings and it always worries me a little bit.

But in terms of applications, um to be honest, I think there's something I really think about for artificial intelligence and and the way we're going with this is

Is it?

Do we want the perceiv the perceptual system of these machines, of these algorithms to be better than human, more human than human?

Or do we want it to be human like?

Like do we want you know on the Star Trek Data, the human like robot.

If do we want the robots to be human like on some level?

And I'm not talking about making wrong inferences through stereotypes.

I'm talking about, for example, estimating social information, perceiving continuous variation, like accent, in the way that humans do.

Because

So, for example, hearing that someone has an accent is is useful information for humans and would be useful information for like AI systems, humanoid robots.

I'll talk about things in the future, of course.

But for me, the next level of like the devices that we communicate with in human-to-spee, human computer interaction is not can we map the acoustics to words, to letters in a

dictionary.

Because to be honest, this is a

Well we already kinda solved it.

It's obviously a much simpler problem than can we map the acoustics to the continuous sensory change of vowel quality, of timbre that allows people to identify age, gender,

dialect, um, all kinds of things.

that and that requires a a much more gradient, sophisticated system than what's required for mapping acoustics to

uh sort of like text entries.

And so I know I'm talking about something that's very far away, but I think that if we are interested in understanding the information that humans get out, for better or worse,

again, so right or wrong, if we want to understand what humans are perceiving, the inferences that humans are making, and especially if we want to sort of integrate this

into any kind of human computer interaction system.

Um

then we need to consider the sort of things that can only be well, that I think that maybe can be more directly addressed through perceptual experiments and and perceptual modeling.

You know what's I I now I don't know anything but you mentioned qu you sort of mentioned in the brain earlier.

I just want to make clear I don't know anything about about the brain or what I'm making infer b inferences based on behavioral data mm-hmm give about acoustic information.

And so the brain understanding the brain is obviously where all of this is going, but

at least as it currently stands, the technology that we have to if to observe the brain can't really get at the sort of the sorts of questions that I'm interested in.

At least not in a in a direct way that is obvious to

Alex Andorra: Right.

Yeah, yeah, yeah.

Yeah.

Uh hope hopefully you'll get there soon though.

That that'd be great.

and so we talked a bit about vowels, actually, and you actually have a paper about vowels.

Um it's called Vowel Normalization as Perceptual Constancy, and it compares what listeners do to the way we understand an object's true size at different distances.

So can you unpack that for us because that sounds very promising.

Santiago Barreda: Okay, so uh I would I've had an example where I was gonna put my hand closer and further from the camera, but since I have these camera focus issues, I don't think I'm gonna do

that.

But here's what I would say.

When an object moves away from you in in the visual domain, the i the size on your retina gets smaller, so cars drive away from you.

You can do this with your hand in front of your face.

Uh but you're you know your brain knows that the object is not shrinking.

I mean this is a

Strange thing to even point out because it's so obvious.

So your brain understands that shrinking person is actually moving away from you.

So it's reinterpreting the difference in the signal as contextual information.

Okay.

So my claim is that this is exactly what's going on in normalization.

You'll notice that a distance commitment in visual perception commits you to size judgment.

What I mean by that is, here's my example.

Suppose that you have a picture of a car and you're asked, is this a toy car or is it a real car?

Well, it depends on the distance to the observer.

If you think it's a toy car, it has to be close to the camera, then it's big.

And if it's a real car, it has to be far from the camera.

There is no real car close to the camera.

You would be looking at the door.

So the two judgments are inherently related.

And and this is what allows us to solve what seems like a circular problem because

If if something it could be just as big but further, there's an infinite number of solutions to this and it's a circular problem.

How does the how do we resolve it so easily?

And I mean part of the way we resolve this is using contextual information in the visual system.

The well known Ponzo illusion, people can Google that, the Ponzo illusion that shows objects of the same size with a background of converging lines, and the objects seem to

greatly change in size.

Um so my claim is

This is what's going on in speech.

It it goes like this.

The vocal tract, the space from the vocal folds in the larynx to the lips are just a tube, like a wind instrument.

When you take a tube like this, you make it bigger or smaller, but you keep maintain its shape, the spectrum, the spectral envelope of the sound shrinks and contracts in the same

way as visual objects do, uh weirdly.

So in what's called uniform scaling.

And the spectrum is basically just a way to mathematically represent uh sounds.

I mean I don't wanna, you know, get too bogged down in the details, but the spectrum determines the timbre of a sound, the feel of a sound.

What makes sounds sound different basically?

Imagine I say a ye those sound different because they have different spectral envelopes, different spectra.

So it hap it just so happens.

Now when you take s a vocal tract and you make it bigger or smaller, you change the spectrum, you shrink it or make it bigger.

So I'm gonna try to do some impressions, but I'm kinda bad at this.

But when you have a short vocal tract that sounds helium heliumies, it's it sounds like this.

I'm trying to maybe my voice is strained.

uh keep in mind that's not high pitch.

High pitch is like this.

That's high pitch.

This is not uh

High pitch I maybe just sound short.

And then long vocal tract is like sounds like slow motion.

Like sounds a long vocal tract.

Because the slow motion slowers all the frequencies, and so it ends up sounding like a uh a much bigger person.

So we have this relationship between vocal tract length and size.

Vocal tract length, because it's the distance from the vocal folds to lips, very very predictably with size.

mean you might not be surprised to learn, but a seven foot adult has a much longer vocal tract than than a three foot child.

And so when people sound tiny, sound small, like like a little voice.

A little voice.

That's a short vocal tract.

And so people use this vocal tract information.

If you hear Shaquille O'Neal and think he sounds huge, and you hear a very short person think they sound small, you're picking up on the spectral information that informs

your size judgment.

And so here's here's that where I tie it together.

In the visual domain, you see smaller things and you know, you say, I know that's not tiny, that's just far.

In the acoustic or auditory domain, you hear someone say, ah, and you hear someone say, and you hear someone say, ah, and you say, that's the same thing.

It's just different sizes.

And so you simultaneous you simultaneously estimate the shape of the sound, which is the same for all these ahs, and you estimate the scaling of the sound is related to the size

of the human.

And the interesting thing is this all this ties together animal communication and uh human speech perception because this is exactly the uh method by which animals convey size

information to each other through their calls.

And there's lot of research on this on deer, koalas, dogs, a crocodile like all sorts of animals that are very separated evolutionarily, uh, use this spectral information and

behave as if they were taking size information from this.

So basically what I'm saying is speech perception is piggybacked onto size communication, which has likely existed in humans for hundreds of of course humans didn't exist, but

hundreds of millions of years just based on the the observation that this exists in all of these very different animal groups.

Unless they all arrived at this ability uh after and independently of course.

Alex Andorra: Hm.

yeah, that's incredible.

And so is that related like is this mainly related to survival?

Like just because there is a very big correlation, very high correlation usually between size and threat?

Santiago Barreda: Yeah, yeah, I mean that's exactly one way I got in trouble once at a talk because some I talked about the results in dogs and someone asked, How do I know what how do you know

what the dogs are thinking?

They asked.

So I was typically, I don't know what the dogs are thinking, I don't know what the deer are thinking.

I'm just guessing based on the results of the in the experiments.

So, for example, there is a study involving red deer that they

this same red deer call and scale the spectrum up or down to simulate our larger, smaller deer.

And they played both calls from sp uh speakers hidden in in the woods and they observed female deer which which speakers they would go to and they would routinely go to the one

that sounded bigger.

So the calls tell uh potential mates this is where the big person is and they also tell potential people that they want to fight.

So so does the deer sort of know that the other deer is bigger?

You know, maybe they d they interpret the information as fear, you know.

Maybe they're just hearing Alex Andorra: so but so I I

don't need to know what the what the animal is thinking.

You just need to observe what they do.

Like you can have a prediction.

Santiago Barreda: Right, so I have a tendency to anthropomorphize him.

I think I might have presented as the deer knows who's bigger.

So they they maybe not don't know, as you said, you don't you d they don't need to specifically feel it is fear, but for all kinds of reasons, just like it's useful in the

visual domain to see who's big and scary.

We know which dog's bark is the big scare dog, and we like if your ball goes over the yard, you know which yard you should not hop into and which is like the little yappy dog

that's no threat.

Alex Andorra: Yes.

Yeah, yeah, exactly.

I mean, this is like instinctive.

Uh and that's funny because like the big dog, yeah, it will scare me, but the small dog it will annoy me.

So it's like there's no way to win.

Santiago Barreda: Yeah, yeah.

And and same with, you know, I mean, fundamentally we're animals too.

And if you are imagine a situation where you're the phone and uh you get into some argument with somebody and you could potentially make a ge guess based on the voice, how

much of a threat this person is likely to be to you if they get all mad, how scary is this person?

Of course, this might be wrong, like I was saying, like famously, you know, Mike Tyson with a slightly high pitched voice.

And s doesn't sound very scary, but he's very scary person, was certainly as a young man.

And so this is an example sort of of the mistakes we can make, but uh but certainly for animals and for humans, there's all kinds of very important survival type reasons why size

communication um and misleading with size.

So that's something that is also interesting in in animal communication.

Because size communication is so important.

All the time animals spend fooling each other into seeming bigger than they are.

Alex Andorra: Yeah.

Yeah yeah.

Yeah, it makes sense.

Do you have do you have examples of that?

Of like animals who are actually tend to be prey but they try to sound more you know, bigger and therefore more threatening than they are?

Santiago Barreda: I mean, I don't know if this is w w I can't speak in all to if this is what caused it, but I do know that koalas have like this like their vocal tract is like three times the size

of their body or something and it because it wraps and they make this insanely low vo low frequency sound that doesn't really match their body.

I I'm pretty sure that's koalas, but I don't I don't know if that is about misleading potential predators, but I'm not super familiar with this literature

But I do know there's a lot it's called um honesty, I think, honestly in in size signaling.

And it's something uh I think uh David Reby, Reby, who works in France, does a lot of this research at in the deer work is his work.

And um he's very interested in communication of size, um cheating, so lying, basically pretending you're bigger so people will stay away from you and out of their territory.

But also catching cheaters and the ability to catch cheaters.

That's a very interesting idea about the development of speech that I'll just throw out there.

Um I can't point to any specific paper, but I had a conversation with him and he mentioned this to me.

Um that he thinks speech evolved from ch cheating and catching cheaters.

He thinks that perhaps um

People started produce being able to produce different vowels because some vowels sound bigger than others, like E sounds small and O sounds big.

And then because humans started producing different sorts of sounds to fool each other, we honed our ability to discriminate those sounds in order to detect the cheaters.

Like I know this old guy's producing an O and he's trying to fool me.

So, you know, I obviously this is just an idea, but it's uh I think it just points to the

interesting relationship between speech and animal communication.

And I just wanna highlight that.

I I find all this really interesting and often linguistics I think forgets that it's related to animal communication, you know.

I would go as far to say it is animal communication.

Alex Andorra: Yeah, yeah, yeah, yeah.

Yeah, I'd never thought about that, but yeah, that I mean that that does make sense to me.

Uh and also something I'm very curious about that may be completely outside of your area of expertise, but and so feel free to tell me.

But you know, I I love um studying accents, you know, and so I really love I'm very fascinated by accents because I find this just

So fun, if I had to model accents, I would definitely use spatial Gaussian processes because from what I've I'm seeing, egg accents are very geographically dependent, but they

are not binary.

They are very continuous.

And so like if you go to France, for instance, where I'm from, uh it's really funny because like the accent really if you go to the northeast, it's gonna be much closer to

German.

And and then as you go closer to Switzerland, it's gonna be of obviously very close to French, Swiss.

and then as you go to the south, you've got that southern accent that's basically very influenced by these uh languages that they had in the south which are related also which

are close to to Spanish, to Italian, and and you can see that in every country.

Like if you go Argentina is another

He's like my second country now and in Argentina it's really like that.

Like the the accent from Buenos Aires is very close to the the accent from Uruguay.

And as you go to the north it changes to the south it changes and as you go to the north, especially north uh west, it gets closer and closer to Colombian accent, Bolivian accent,

and and these accents are all very close.

Um so I found that fascinating and that that

Brings me to my question which is do we have any work and any ability to actually kind of predict uh the way languages evolved and and then will evolve?

Like in and my question is like, can we actually have a decent idea of what the map of languages could be in a f like, you know, in a century or or maybe

Maybe longer because we were able to retrace how we ended up with the modern English, modern French, blah blah blah.

Or is that completely out there?

Santiago Barreda: Well, you know, that's a very active area of research in linguistics, accounting for why changes happen, how they spread, why some changes spread, why some don't, um, where we

might be going in the future, understanding and delimiting change.

so this is a a very active area of research.

Admittedly it's not one I know very much about, but I will say we again

I think this is a recurring theme in in linguistics.

We know kind of n nothing a little bit, like in terms of like being very sure.

There's a lot of ideas, some of them are very reasonable, but we're not really close to where we could say, uh, this changed because of these reasons.

Uh other than, you know, maybe like there might be obvious cases like the Norman conquest of England caused some changes, you know, French speakers arrived.

These th some of these things are very obvious, but I just mean the step of you're saying continuous gradient acoustic variation across a continuous geographical area.

I don't know of any work that sort of can explain this in in the general case, and certainly predicting the future changes is also, I think, something that's very poorly um

understood.

You know, my take on this is

I think language language change is a little bit like a chaotic system.

And so I mean like deterministic locally or or you know reasonable but but chaotic in like I think I think you know talking about the future of the English language is a little bit

like trying to predict.

Well the weather in the you know, where will it rain?

Yeah, exactly, exactly, yeah, exactly.

So um that's kind of how I see it.

I don't really see the failures a as a failure of theory.

It's just like yeah, w we're not really there yet.

Uh but we know a lot about the sorts of things that matter, the sorts of things that don't.

Uh we have good descriptions of previous things.

Um but yeah, I think we're we're pretty far off.

But what what we're really good at on the positive side is is quantifying these things.

So there's standard ways to measure vowel acoustics, to do uh what are called vowel plots and

When you plot the where the vowels fall on like a frequency space for a dialect and then another dialect, um you this is highly predictive of uh cross language perception,

cross-dialect perception.

And you can also do things like cross generational plots.

And you can see how within a community the s language system might be changing as a function of time, like relative time.

You can also get recordings across time.

And so with these tools, potentially

you can make inferences like, you know, ah is becoming more like and if it's continuous, you could say, well, if it keeps along the trajectory, it will go th here.

But then maybe it stops.

Alex Andorra: Yeah.

Santiago Barreda: You and you don't know why.

Alex Andorra: Yeah.

Santiago Barreda: So

Alex Andorra: Mm-hmm.

Yeah.

Yeah.

Yeah.

No, I mean Yeah, I c I mean that fits also my prior that this is much more of a chaotic system where it's more an accumulation of small, somewhat random changes that just then

end up making big changes, but it's super hard to predict by definition.

Um and at the same time, in retrospect you can definitely see also, yeah, why some words, you know, evolved e the way

They did and like, you know, for instance, I don't know for sure, but uh in in English, you know, the word donut I see it also written doughnut with like D-O-U-G-H.

I'm guessing this is the way it used to be mostly mostly written because it makes much more sense.

But now it's just D-O-N-U-T, right?

And so

Yeah, like these kind of things where we can see it it evolved, but it's a more hard to predict, you know, but like uh so plus you have some languages which are more conservative

than others.

Um like French.

Very famously we're very conservative and and we haven't have an academia to actually change words.

So this is it's very weird.

But yeah, like that that for sure changed like make changes much slower.

Um

But yeah, like I I really love thinking about the these these topics because I find that really fascinating and and you know, for instance the fact that and how also what I found

super uh impressive is how languages die, which is also very interesting to me.

Like how can Latin, for instance, have died?

Because small you know, dialects in France or Italy, you

Can imagine that they died because well the state, like the central state, was very much against them and it was not the language of the winners.

But Latin was the language of one of the biggest empires, and it still died.

So to me that's that's really interesting to understand that.

It's like literally nobody speaks Latin anymore.

Why is that?

How is that possible?

Santiago Barreda: Well, you know, the thing is Latin I mean, in a sense it's dead because it certainly doesn't exist in anything like what it used to be, but it kind of became Italian and

Spanish and French.

Alex Andorra: Yeah.

Santiago Barreda: It's a little bit like, you know, dinosaurs, we we f we fry them and eat them.

They're chickens now.

Like a little bit like that.

So uh the it is interesting.

Alex Andorra: Mm-hmm.

Santiago Barreda: Wha some changes some languages change so much that they become unrecognizable and then

When do you consider that a language is a legitimate different language?

Uh I mean these are all complicated issues in linguistics that are certainly not settled.

Because even for example, uh if we live in California and we're talking about what is the future of California English?

But before we can even discuss that, we can say, well, what is California English?

You know, who is the regular, normal California person?

I live here thirteen years.

I guess maybe I don't count as a real how long do I have to live here?

To be a real California English speaker.

And or and you are you are you a California English speaker?

How long do you have to be here?

There's also a lot of ethnic variation, socioeconomic variation, gendered variation.

So to even talk about change, we need a reference point.

And necessarily that makes decisions about that makes means we have to make decisions about who is a normal speaker.

And so I just be this is like further interesting wrinkles and uh interplay between dialects of English, for example.

An easy example is African American English is uh has a gr tremendous influence on the development of mainstream, well what can we call mainstream American English?

Uh but on the other hand, these are all they influence each other but also evolve on their separate courses.

Alex Andorra: Mm-hmm.

Yeah.

Yeah.

Yeah.

Um makes me think reminds me of like there is often also, you know, these debates about actual French names or, you know, first names of or things like that.

And that's always something where I'm like, what does that even mean?

You know, like my name, for instance, Alexandre is considered a French name, but it's not.

It's it's like it I think it's Russian to begin with.

When did it become a French name?

You know?

I I'm sure when the first people named their child Alexandre of like a few decades ago, people were saying, that's not a French name.

Um so when does that you know, when does that happen?

Where where's the where's the actual boundary?

I don't think there is any boundary.

I think it's a continuous um process, much more like accents and the and the way um and the way languages evolve.

and speaking about animals

Uh speaking I see that your cat was successful in in in in telling you that he needed to be he needed to be on the show.

Well done.

Talking about that, yeah.

Cat evolution is also something that's extremely extremely fascinating.

I I always love learning about that.

Very interesting animal.

Um awesome.

Well

It's starting to to be late here Santiago, so I don't wanna take too much of your time.

But I do wanna ask you if yeah, what what are you you do a lot of things and you're a very curious person, so yeah, looking forward, what are you focusing on for the coming month

and is there anything you're trying to learn actually right now that uh that your that's one of your focuses?

Santiago Barreda: Um well I'm still sort of working mostly on this normalization issue.

one of the things I'm most interested in i is developing a working on developing a Bayesian model.

Well I would say I'm developing it further because it was proposed by Terry Nearey and Peter Assmann in two thousand seven.

It's called the probabilistic s sliding template model.

And uh

It's twenty years old, but no one's ever really read the paper because it's very dense, it's like a little bit indecipherable.

So la two last year I think Florian Jaeger and I published a sort of reintroduction of of this paper.

And it's a Bayesian model of joint inference of indexical information and speaker information that can simultaneously guess what the person is saying and indexical

characteristics basically of the speaker.

Using the same sort of social knowledge that we hypothesize that that listeners have.

But the really good thing f in my opinion is that the model doesn't just guess categories.

Because like I said earlier, that's kind of the easy part.

Um it tries to guess the continuous phonetic information.

So like where it falls in this continuous vowel quality space.

So it makes all three sort of judgments or more potentially simultaneously.

And I'm really interested in applying this model to uh listener data to see uh if it works, what needs to be changed, if fundamentally the the approach is as useful as it

seems like it's gonna be.

So that's one thing I'm working on, just refining this model, applying it to listener data, so collecting perceptual judgments.

And I'm one thing I'm trying to learn is more about wave acoustics.

really trying to wrap my head around all this very frustrating stuff related to impedance matching and reflections and blah blah blah when you model the vocal tract as a series of

concatenated uh short tubes to understand to better understand the relationship between vocal tract shape and uh the acoustic signal.

And applying that to sort of ex um

A a better I'm trying to just arrive at I'm trying to come up with more intuitive explanations of of vowel acoustics because I'm in the same way that the stats book is

trying to just apply to intuition and not use math to speak for itself, a lot of the explanations in linguistics of vowel acoustics sort of allow the wave physics equations to

speak for themselves in a way that is not helpful to people that are not engineers or very, very strong at at math.

Um and so I'm trying to learn that and part of the reason I'm trying to learn that

is uh so that I can I'm working on in a long term on a book on this, but I'm also continuing to make software that carries out uh acoustic analyses and helps people do

better sort of statistical modeling of production data by uh developing a GUI version in Python of that FastTrack of this thing I wrote for Praat that helps people

carry out high quality formant analyses because the way that this is often done in linguistics is just

Digital signal processing is just applying a model to speech, actually.

So it's just the best guess of the parameters given the model you set up.

And people are often applying wrong models using default analysis settings to their speech and getting really bad measurements.

And FastTrack, again, actually the original algorithm was developed by Terry Nearey like tw twenty, thirty years ago, but he just never publicized the software.

And um it just helps people applies statistical techniques.

Very simple ones, but does something to try to find the best analysis parameters for a piece of speech recording instead of just assuming that the default is the best.

And so I wrote this in Praat because everyone is a standard linguistic software that even undergraduates soon.

So I by that I just mean people with very little research experience can use.

And now working on a better version in Python because

Faster, it's better, etcetera.

But then people need to be willing to install it.

So the barrier for entry is just slightly higher than Praat's.

Alex Andorra: Yeah.

Well, that sounds that sounds like fun.

Uh yeah, I know you've you've you've worked on these on these tools for phoneticians.

Right.

Uh so that that one for for Praat, phonTools, FastTrack.

Do you wanna touch on on that very quickly before I close this out, actually?

Santiago Barreda: I I might more or less just I I did I wrote phonTools as a grad student just to sort of see if I could and to learn.

Kind of applying the same principles from the Bayesian stats book.

Instead of reading like for example, early on the first thing I wrote was a spectrogram because I wanted color spectrogram.

So spectrogram is a representation of the frequency content in the sound as a function of time.

And I read the description of the math and I wanted to make sure I understood it, so I wrote a function that does it.

And that's kind of what I did for that R package.

I just wrote a bunch of functions that do these things I was trying to learn about.

And that was the purpose of that.

But it seems like people use it.

I'm I never hear from users or readers by the way, so I can't tell if anyone reads anything or uses everything anything.

I uh only citations is what lets me know that at least somebody is using this stuff.

Alex Andorra: Yeah.

Well if somebody is earlier is listening, please uh please let Santiago know.

Yeah.

Santiago Barreda: Even if you hate it.

Alex Andorra: Yeah, yeah, exactly.

His contact is in the show notes.

Uh so you know.

You'll have his website and and contact him.

That's always very helpful.

I completely agree as somebody also who writes uh public tools, um, to have to have feedback and and advice from users.

Uh

Santiago Barreda: Yeah, I never hear from anybody.

So please I would love feedback but just remember I just keep in mind I'm very sensitive.

So I

Alex Andorra: Yes.

also well Santiago anything you you wanna add uh that you wanted to touch on and and didn't yet before I ask you the last two questions?

Santiago Barreda: Um mm mm mm mm.

I don't think so.

Yeah.

Alex Andorra: Awesome.

I think we did cover a lot of ground, so I'm happy about that.

Santiago Barreda: The o here's the only thing I would say.

I think that I would have the reason I wrote this book is basically because brms exists.

It's an R package that basically writes Bayesian models for you instead using lmer-type syntax.

I think that I would have never written the book for using JAGS.

I didn't think it was realistic for people to write very these very complex JAGS models.

It would take me very long time.

It would fail.

I would find out a three days later, redo a little typo.

Um so what I would say is

Using brms, the barrier for entry in for Bayesianism is so low that there's really no reason to not use it other than potentially the slowness of the models.

That's a different issue and that's can be avoided sometimes.

And people's main sticking point is stuff like worrying about priors or things like that.

I think these concerns are overblown in, I'll say, in linguistics.

The type of data we have, the priors are gonna have minimal impact on your results, and you can even check if they do or don't.

I would encourage people working in linguistics, uh, perceptual data to give Bayesian stats a shot.

What I say a little bit tongue in cheek is you can make your Bayesian model as terrible as any frequentist model.

And by that I mean you can make it analogous if what you want is something analogous with like uniform priors and and minimal constraint things like that.

Uh but you can make them much better and you can get really, really useful information about what are called random effects in your data, which

When your random effects are listeners, what you're talking about is listener variation in the speech community.

So these things are actually tremendously useful to know about.

And so in speech research specifically, the information you get and the flexibility of multi-level models, uh, in handling outliers, in head handling sc heteroscedasticity,

ordinal models, multinomial models, these things that are really end up being necessary for speech research.

Um

I really think once people give it a shot they'll n see immediately the benefits and

I just hope people will uh be open minded and and and uh give a Bayesian approach a shot.

And I think that they will be very happy.

Alex Andorra: Yeah, yeah, yeah.

I mean, love that.

Love that call to action, Santiago.

Obviously I resonate with all of that.

Um actually I gave a talk recently at Princeton to the psychology department where I went exactly through these topics in the presentation and the workshop.

I'll put the link to the GitHub repo.

That I set up for that.

And it's actually a notebook where yeah, people can go and see how you would do a mixed effect model.

Because I know it's used a lot in psychology, especially the the lmer, lmer uh flavor of these models.

Where I show you exactly how to do that in the Bayesian workflow with PyMC, so in Python.

That way you have brms, you have Python, you can choose whatever you prefer.

And um

And you can see the effect of the priors.

I see actually I show actually different different priors.

And and I show that actually having a wide prior doesn't mean it's a good prior.

Um so yeah, to to echo exactly what you said.

and now it's so much easier to to start with these tools, not only because we have supercomputers in our hands and and our

Personal computers now are so powerful.

and thanks to um open source uh software like brms, Stan, PyMC, we can actually sample MCMC in a matter of seconds.

So that's definitely a game changer.

And also now you can get some assistance if you do it well from uh AI agents.

So I show you that also in the in the talk.

So if you're folks, if you're interested, I'll

I'll put the link in the show notes to the GitHub repo.

I think I'll also release the at least the talk, not the workshop, but the talk as a as an episode.

So probably after yours uh at Santiago, probably I'll do that so that people have that.

And prol and if I can I'll release the the workshop also, but just on YouTube because doesn't work well as a as a podcast episode.

Hopefully that's useful.

Um and of course, huge shoutout to

all the community who makes that possible.

PyMC core developers, Stan core developers, and brms, who was created by Paul Bürkner.

He was on the show a very long time ago.

That was episode 35.

But I'd put it in the show notes after if you want to listen to that related episode.

But also, Paul was actually just back on the podcast, just before your episode, episode 167.

Santiago, that was the live special episode at StanCon and Paul Bürkner was uh a recurring guest there, so I encourage people to to listen to that because he always does amazing

work and and influential work.

Sorry.

Yeah.

Alex Andorra: Yeah?

Well, me too.

Yeah.

Paul, if you're listening, uh thank you so much.

Vielen Dank.

And well, Santiago, to before I let you go and and and we go have dinner, I have to ask you about the the last two questions.

So first one, if you had unlimited let's do that again.

If you had unlimited time and resources, which problem would you try to solve?

Santiago Barreda: Well

You know, taking the question literally, I should definitely say something like cure cancer or something, right?

Because what my answer about my research would be a terrible answer if this was real, because definitely need to do something more important.

Solve r you know, cure reduce world hunger or something.

But if I'm gonna keep it to my research, which sort of, you know, limits the the sort of interesting things I might say.

I it's something I'm currently working on, but it's very time consuming.

Do this properly, but it would be very useful to have a very large database of speech from pe speakers of a dialect, different sorts of speakers, and have listeners of that dialect

listen to this speech and rate and tell you what vowel they think they're hearing, rate the speaker characteristics and so on.

And what I'm describing is obviously it's just a simple experiment that I can do at any time.

But I'm talking about a lot of speech produced by a lot of people and people

And listeners potentially listening to hundreds or thousands of these stimulus.

So I'm talking about a very large amount of data to use to sort of test different models of the inferences that people make.

Especially the elusive, continuous phonetic judgments of quality.

So within category variation.

So you see two shades of red and you c call them red, but you clearly see the differences.

So understanding

that's that level of detail for real speech requires a a very large amount of of data collected.

And ideally in a lab is the problem because online studies you always worry about is this really the dialect, are they really s listening, et cetera, et cetera.

So I would really love to get my hands on this data.

This is something I'm working on right now, but doing this with a large number of speakers and a large number of stimuli.

Um

Is it is slow and time consuming for somebody who's like, you know, doesn't have a big lab uh working for them sort of thing.

Alex Andorra: Hm.

Nice.

Yeah.

Yeah.

Santiago Barreda: It is said publicly, but I'm sorry, I'm sorry.

I would release it publicly by the way.

So my d I would love to create this data for people to use, not like uh just for me.

Alex Andorra: Mm-hmm.

Yeah.

Yeah.

That's awesome.

Very much very much in character, so so I love that.

Um and second question, if you could have dinner with any great scientific mind, dead, alive, or fictional, who would it be?

Santiago Barreda: Okay.

I need I really need to take again I need to need to take advantage of this opportunity because there's really a lot I could do with this.

So I my answer is kind of cheating a little bit.

But I think it's within the bounds of of of cr of okay.

Because I think I would like to talk to Socrates and see.

People might think, Well he's not a scientist, but this is my defense.

He's not a scientist in the modern sense, of course.

But given the time he lived in, the sorts of

his approach to knowledge gathering.

I know he was interested in more on like, you ethical or or metaphysical questions, but the the way he approached knowledge gather knowledge gathering, his uh s his uh the way he

used logic and counter arguments, the way he was really seeking truth, I would say he had a scientific mind for sure.

And so I would love to talk to Socrates for two reasons.

First, most of what we know about him is other people like hearsay.

So I would love to know

What he really said is like how much of Plato just putting words in his mouth.

So this would be a very interesting conversation, I think.

But also I would really love his take on the current state of knowledge in the world and the current state of society.

And uh if he could like control himself, sort of and not like and not agree to argue too much or get too mad at things, I think.

And he we could understand each other, I think it would be uh a really interesting

opportunity to get the perspective of a a very, very intelligent and wise person uh on w how things have developed and what we know, you know, imagine him thinking about atoms and

magnetism, things like that.

Alex Andorra: Mm.

Mm-hmm.

Yeah.

I'll definitely allow it.

I completely agree.

Uh the Socratic method and way of being curious about the world definitely qualifies as uh a scientific um thinking.

So uh I would say it's probably one of the basis of of scientific uh reasoning.

So definitely.

Um definitely allowed.

Uh thank you so much, Santiago.

Thank you.

I was great to have you on the show.

Um

I feel like we could we could keep talking, but that's a good a good time to stop.

yeah, thank you so much for taking the time and being on the show.

for having

This has been another episode of Learning Bayesian Statistics.

Be sure to rate, review, and follow the show on your favorite podcatcher and visit learnbayesstats.com for more resources about today's topics, as well as access to more

episodes to help you reach true Bayesian state of mind.

That's learnbayesstats.com.

Our theme music is Good Bayesian by Baba Brinkman.

feat. MC Lars and Mega Ran.

Check out his awesome work at bababrinkman.com.

I'm your host.

Alex Andorra.

You can follow me on Twitter at Alex underscore Andorra like the country.

You can support the show and unlock exclusive benefits by visiting patreon.com slash learnbayesstats.

Thank you so much for listening and for your support.

You're truly a good Bayesian and change your predictions after taking information.

And if you're thinking I'll be less than amazing, less adjust those expectations.

Let me show you how to be a good Bayesian.

Change calculations after taking fresh data.

Those predictions that your brain is making.

Let's get them on a solid foundation.