The cocktail party effect

Chris -   Hello.  Welcome to the eLife podcast episode number 3, produced by the Naked Scientists.  I'm Chris Smith.  In this edition, how skin cells were turned into retinal cells to discover the cause of a patient's blindness, improving vaccines by studying the genes that are switched on when a student gets a flu jab, and the stem cells that might be the Achilles heel for one of the worst human parasites.  First though, to the question of how our system of hearing, despite being constantly assailed by a barrage of sounds from a range of different sources can pay attention just to those sounds that are meaningful - for instance, just one person's voice in a crowded, noisy room.  Sundeep Teki from UCL has discovered that the brain is highly tuned to spot when certain frequencies start and stop at the same time.  This is how it knows which ones to listen to.

Sundeep -   In a realistic environment, we have lots of sounds which have overlapping frequencies, the start and stop at different times.  So, what we try to do in our study so far is to come up with a signal which more accurately represents what we really hear every day in our life.  So, instead of our previous work which was just based on two frequencies which were alternating in time, we came up with a more complicated stimulus which has several frequencies in it.

Chris -   I've got a sample of that, so we could just have a listen to this (sound recording)...

Sundeep -   In a way, it's a bit like noise.  So, what we did within this sound was to make a few frequencies, start and stop at the same time and what happens is that that particular segment of sound kind of pops out against the background.

Chris -   Well, let's have a go.  (sound recording).  And there's a very obvious stand out difference, isn't there?

Sundeep -   Exactly, so there is this pop up effect and listeners can get it quite quickly.  We didn't have to train them too hard in order to get that.

Chris -   Well, what does this show?  Why is this important?

Sundeep -   Right.  So, the pop up effect suggests that the brain has some kind of a mechanism which is able to extract this pattern that we've imposed based on a few frequencies that start and stop at the same time.  So, we think the brain is sensitive to these correlations and it tracks the time on the signal.  And if certain frequencies start and stop at the same time, it registered that there was something here and it might lead us to suggest that these sounds might belong to one particular object.  So, we know that in this background of multiple sounds, we heard something salient which popped out.

Chris -   Does this mean then when I'm having a conversation with someone and they're speaking to me, that I am at the same time is listening to what they're saying, I am also running ahead in the conversation, almost predicting what I think they're going to say and then my brain is listening out for that predicted regularity of pattern, to then extract that information from the sound?

Sundeep -   Indeed.  Our voice is very rich signal, not only frequency and time information, but also information about your vocal tract.  So, after a few seconds, I think the brain kind of forms a template of your voice and it's building up with more and more information.  The model prediction becomes better and better.

Chris -   So, when you've got your subjects in the laboratory and you're playing them these sounds, what sorts of measurements are you making?  How are you testing the person's reaction to those sounds and the differences?

Sundeep -   So, what we did is, we invite listeners to a sound proof booth.  Sounds are played which have different properties.  So, we varied two parameters, the number of frequencies which became synchronous and the time for which it became synchronous.  And the task of the listener was who wants to press a button as soon as they heard something pop out.  So in each block, half of the sounds had a target in them and half of the sounds did not have a target in them.  And basically, what we found is that they are extremely good at detecting these targets even with very little practice.  And the performance improves if the number of frequencies that are synchronous are more and more, and the time for which they are synchronous are more, suggesting that the brain is kind of integrating information from both the frequency and the time domain to be able to detect the target.

Chris -   So, armed with those observations, what are you concluding and what will be the logical next step?

Sundeep -   Apart from doing this kind of behavioural testing, we also did a modelling analysis in our study.  The main hypothesis of this model is that the brain is indeed looking at the time information and trying to calculate correlations between different frequencies.  So, if there is a strong correlation in time for certain frequencies, then it might belong to one object and they might stand out from the rest of the background.  Our data which is based on a signal which is very realistic supports this model which is known as the tempo-coherence model.  We think that this is the way forward.

Chris -   Because practically speaking, where we would really like to go with this apart from understanding how we work is to endow our computer with the same abilities that we have because I'm sure everyone who's ever spoken to a computer down a telephone and tried to tell it what you want, the problem is obvious.

Sundeep -   Yeah, I know what you're talking about.  So, speech recognition is still very intense topic of research.  Results from fundamental neuroscience studies like this one may help inform the engineers to come up with more sophisticated algorithms that they can implement in their devices.  So hopefully, we hope that this might lead to some kind of a practical benefit in hearing aid design or machine listening devices, and so on.

Chris -   Sundeep Teki.  He's at the Wellcome Trust Neuroimaging Centre at UCL.