Periodic Reporting for period 1 - VIOLA (Variability's Influence on Language Acquisition)
Reporting period: 2015-09-01 to 2017-08-31
Some babies spend most of their waking time with only one person; others hear speech from many different speakers. Even if the two babies heard the very same words, their experiences would not be equivalent. In fact, although for us adults “dog” said by a girl and “dog” said by a grandfather mean the same, careful analysis reveals that words sound very different depending on who speaks them. So are there effects on how babies learn language depending on how many people speak to them? When we started this project, the answer was far from obvious. Many previous researchers had tried to look at this question by doing laboratory experiments. For instance, one experimenter played a recording with many repetitions of a made-up word spoken alone (“beece, beece, beece”) and then passages using the word or not, to test for simple recognition. Another taught babies two made-up similar words (a “beece” was a round yellow toy, and a “peece” was a square red one) using only one speaker’s voice or several speakers’ voices. It turns out that results from these studies were really variable, with babies hearing many different voices performing better or worse than babies hearing one or a few voices. Results probably differed because the methods varied -- but what we realized is that a real baby, in the real world, needs to do all of these things: recognize words, and learn to map words to objects in the world. It turned out that nobody had thought about our key question: Does number of talkers matter in the real world? We knew this question was difficult to answer, so we used every relevant method to try to answer it, as we explain in the next section.
Is language acquisition different depending on how many people speak a child? We used three methods to answer this question. First, we ran three laboratory experiments with infants, and a fourth experiment is ongoing. Overall, we have found relatively little evidence that the number of speakers babies are routinely exposed to impacts their language development. This might entail that infants have ways of dealing with talker variability beyond what we had given our computer models, either as part of their innate language skills or because something in their experience allows them to compensate for that variation.
Our second method is called a “meta-analysis”, and it consists of a detailed, quantitative review of all previous experiments on one topic, which allows us to combine all evidence into one metric. We carried out three of them, on slightly different questions, which helped us identify good methods to use in our experimental work. In the process, we also made the same data available to the broader community (see http://metalab.stanford.edu). These meta-analyses helped design more informative infant experiments.
Finally, we used computational models to check how speech processing was affected “in theory” when there were multiple talkers. Using a computational model, rather than human participants, is very informative: Since we design the model ourselves, we know exactly how it perceives sounds, and we can thus check what an unbiased ear may do with a single-talker or multi-talker experience. Another advantage is that we can do many more tests than would be feasible with actual human listeners. In a nutshell, the computational models showed us that at the level of the human inner ear, we should find it a bit harder to distinguish sounds when they are spoken by different people than when they are spoken by a single person. Another, rather incidental, result was that it mattered a lot which sound pair is compared, and this is important because studies with adults and infants often cannot compare more than a handful of sound pairs. The results from such studies might not reflect reality across the board, because we find that some sound pairs are always easy to distinguish while others are always hard, both extremes show little influence of the differences between talkers. Sound pairs that are a bit difficult become harder to distinguish when multiple voices are involved. A second modeling study was conducted on data from meta-analyses to find out whether and how knowing words and knowing sounds interacts over the first years of life.
The results have been disseminated widely: two papers that have been accepted for publication in peer-reviewed journals and are available as preprints, three papers are already published in peer-reviewed proceedings; further manuscripts are in preparation. The results have also been disseminated in five invited lectures and at ten international conferences.
Our second method is called a “meta-analysis”, and it consists of a detailed, quantitative review of all previous experiments on one topic, which allows us to combine all evidence into one metric. We carried out three of them, on slightly different questions, which helped us identify good methods to use in our experimental work. In the process, we also made the same data available to the broader community (see http://metalab.stanford.edu). These meta-analyses helped design more informative infant experiments.
Finally, we used computational models to check how speech processing was affected “in theory” when there were multiple talkers. Using a computational model, rather than human participants, is very informative: Since we design the model ourselves, we know exactly how it perceives sounds, and we can thus check what an unbiased ear may do with a single-talker or multi-talker experience. Another advantage is that we can do many more tests than would be feasible with actual human listeners. In a nutshell, the computational models showed us that at the level of the human inner ear, we should find it a bit harder to distinguish sounds when they are spoken by different people than when they are spoken by a single person. Another, rather incidental, result was that it mattered a lot which sound pair is compared, and this is important because studies with adults and infants often cannot compare more than a handful of sound pairs. The results from such studies might not reflect reality across the board, because we find that some sound pairs are always easy to distinguish while others are always hard, both extremes show little influence of the differences between talkers. Sound pairs that are a bit difficult become harder to distinguish when multiple voices are involved. A second modeling study was conducted on data from meta-analyses to find out whether and how knowing words and knowing sounds interacts over the first years of life.
The results have been disseminated widely: two papers that have been accepted for publication in peer-reviewed journals and are available as preprints, three papers are already published in peer-reviewed proceedings; further manuscripts are in preparation. The results have also been disseminated in five invited lectures and at ten international conferences.
The VIOLA project produced a host of new insights that propel theory building in early language acquisition forwards on a number of levels. First, competing accounts were assessed and none could capture all results in a satisfactory way. It seems thus that current models of early language acquisition need to be refined with all available data in mind. Second, implicit assumptions theories are built on have been assessed, with the results favoring accounts that can capture the negative impact of variability on the acoustic signal and that do not rely on complex interactions between linguistic levels for acquisition. Finally, the reliability of extant infant data has been examined and found to be sub-optimal due to underpowered research. This might in part explain the seemingly conflicting results and thus leads to stronger theories when considering which outcomes are reliable and replicable.
We now turn to implications for society. During the project it became clear that the main variable of interest, namely talker variability in daily life, is closely tied to three distinct factors in infants' environment: (1) Does the child attend daycare, and if so which form of daycare (single child or a few children and one caregiver or groups of children and many caregivers); (2) How big is the core household?; (3) How many social interactions take place regularly (visits of friends / family, play groups)? These three factors all shape infants' language input in distinct ways, but we know as yet little about their impact on language development. At the same time, aspects of the first factor is crucially shaped by national and European policies.
Based on the results obtained here, it is possible to tentatively state that all types of caregiving arrangements are similarly beneficial, and that there is no single type that is superior to others. Because the sample tested here is socioeconomically homogeneous (with highly educated mothers), we caution against further conclusions, especially when it comes to multilingual acquisition.
We now turn to implications for society. During the project it became clear that the main variable of interest, namely talker variability in daily life, is closely tied to three distinct factors in infants' environment: (1) Does the child attend daycare, and if so which form of daycare (single child or a few children and one caregiver or groups of children and many caregivers); (2) How big is the core household?; (3) How many social interactions take place regularly (visits of friends / family, play groups)? These three factors all shape infants' language input in distinct ways, but we know as yet little about their impact on language development. At the same time, aspects of the first factor is crucially shaped by national and European policies.
Based on the results obtained here, it is possible to tentatively state that all types of caregiving arrangements are similarly beneficial, and that there is no single type that is superior to others. Because the sample tested here is socioeconomically homogeneous (with highly educated mothers), we caution against further conclusions, especially when it comes to multilingual acquisition.