The Tower of Babble

Māma, mama, mamá, ma, mama, mamã, maa, mama, haha and mamī.

• The word for ‘mama’ in the 10 most widely spoken languages in the world.

When I think about baby language learners, an image that often comes to my mind is of an eight-month-old girl who came to our lab a few years ago. Let’s call her Barbara. Her mother was Swiss French and spoke to Barbara in French. Her father was Czech and spoke to her in Czech. Her parents spoke to each other in German, as that was their most fluent common language. They had recently moved to London, so much of daily life outside the home was in English. They also had a Spanish au pair, whom they encouraged to speak to Barbara in Spanish. On a typical day Barbara heard five different languages. She might not become fluent in all of them, but knowing multiple languages is no problem for a baby. In fact, monolinguals are currently a global minority, with only 40% of the world’s population knowing only one language. Lots of babies grow up multilingual.

We are born with a head start on our mother tongue. Babies have heard a muffled version of it in the womb and picked up a few clues. At just one day old, baby Barbara would prefer the sound of her mother’s voice to that of a stranger. If you filtered her mother’s voice to make it sound muffled and distorted, the way it was in the womb, she would prefer that even more. Studies show newborn French babies prefer the sound of French over Russian. Interestingly, they also prefer English over Japanese but have no preference between English and Dutch (Nazzi, Bertoncini & Mehler, 1998). This suggests that babies are picking up the broad patterns of their language, which lets them notice when they hear a language from the same family.

Babies are born ready to tackle any language on the planet. As an English speaker I have lost the ability to hear the difference between Hindi speech sounds /Ta/ and /ta/ or between the click-like sounds <xa> and <ca> found in Zulu. If I was Japanese, I would have trouble with English /ra/ versus /la/. Newborn babies arrive able to tell all these sounds apart – and more. These sounds are called phonemes and they are the building blocks of words. Spoken language is a messy business but, roughly speaking, phonemes are the consonants and vowels of that language. English has about 24 consonant sounds and 20 vowels (compare the different sound of ‘a’ in the words back, bra, bay, bat, ma and mama). French and German are similar to English but there is dramatic variation. Modern Arabic has just three vowels, while the !Kung language has 141 phonemes, thanks to its use of click consonants. Tonal languages like Mandarin make things interesting by letting changes in pitch alter meaning. Foreigners learning Mandarin often get challenged by the sentence ‘Mā má mǎ mà ma’. The different accents mark rising and falling pitches, and the sentence translates as ‘Mum is bothered by the horse’s scolding – yes?’ This may be nonsense, but it is clear and unambiguous to native speakers. Babies are born able to hear all these contrasts, but by the end of their first year they tune in the sounds of languages they will use (Werker & Tees, 2003).

Speaking is tough. Psycholinguist Peter MacNeilage calls speech an ‘invisible miracle’. He estimates that to speak fluently at full speed we must coordinate 40 different muscles and orchestrate about 225 muscle movements every second (MacNeilage, 2008). It is no wonder babies take around a year to say their first clear word. They need a lot of practice. From around two months of age there are ‘goos’, ‘gahs’ and a whole range of other random squawks and squeaks. The ‘oohs’ are the start of vowels, the vocal sounds created by air flowing evenly through the vocal tract. Changing the position of the tongue in the mouth and the shape of the lips changes the vowel. English has so many vowels because it allows diphthongs, which slide from one vowel sound to another as in ‘wait’ or ‘loud’. Consonants are more complex still. They are created when the airflow is interrupted in some way. These arrive as babbling begins at around four months, but are still being worked on two years later.

Babbling is an essential stepping stone to speech. Like musicians practising endless scales and arpeggios before graduating to melodies, babies are getting into the flow of speech. By six months babbling starts to have recognisable syllables. The first consonants to arrive are b, d, m, n and w, and babies practise them one at a time. Parents eavesdropping on babies put to bed while still wide awake will hear ‘mamamamama’ and ‘dadadadadada’ coming through the baby monitor. The words /ma/ and /pa/ are an interesting special case. They are found with minor variations in hundreds of languages across the globe. Some researchers argue that these words get reinvented many times, as parents pick these sounds out from babies’ limited early repertoire. Linguists Pierre Bancel and Alain de l’Etang argue that /ma/ and /pa/ are too widespread and consistent for this to have happened by chance, and must instead be words inherited from an ancient common ancestor. They are so resilient because the ‘mmmm’ and ‘aaaaah’ sounds are so easy for babies say and because the words are so important (Bancel & de l’Etang, 2013). ‘Mama’ is not always a baby’s first word, but it is one of humanity’s and is undoubtedly one of the oldest words in existence.

Early babble seems monotonous, but even in a song of one note there is a lot for babies to learn. They are learning about changing tempo and stress and about which vocal actions make which sounds. They are also learning to keep going. Talking is like walking – you need to string a whole sequence of steps together and you need a lot of practice. Deaf babies babble in sign language (Petitto & Marentette, 1991). As babbling progresses, babies mix it up, talking a delightful nonsense that sounds like a private language. They string together a wider variety of syllables. Babies’ babbles capture some aspects of their mother tongue, particularly intonation. But their ‘words’ are mostly not recognisable. We can usually tell a French babbler from a British babbler from a Chinese one (Petitto & Marentette, 1991). It is almost as if they are imitating their idea of what they hear around them. Or maybe it is freestyle jazz?

To be fair to babies, even with infant-directed speech, adults can be hard to follow. We do not notice this for our own language, but speakers of foreign languages sound as if they are speaking fast and speaking indistinctly. This is sometimes called the ‘gabbling foreigner illusion’ (Cutler, 2012 ).

Weallspeakfastanddonotleavemanypausesbetweenwordswhenspeaking.

If you look at the spectrograms  of recorded speech, you see that the pauses do not line up nicely with the intervals between words. This makes life difficult if you are trying to learn that language. It is hard to pick out words if you have no idea what they might be or where one ends and another begins.

In the late 1990s three American researchers, Jenny Saffran, Richard Aslin and Elissa Newport, realised this must be a problem for babies and wondered how they solved it. They decided that babies did it with statistics. Suppose you are a baby. You will hear lots of phrases like ‘Youwannajuice?’, ‘Youwannabottle?’ or ‘Youwannacuddle?’ and ‘Areyousleepy?’, ‘Areyouhungry?’, ‘Areyoumummyslittlemonkey?’ Hearing variations like these and you might start to think ‘Youwanna’ and ‘Areyou’ are blocks and that the gaps come after. Combine lots of examples like this together and you start to notice that some syllables clump together, which helps you segment out the words. When someone says ‘Hellobaby’, the baby hears ‘Hello baby’ not ‘Hell obey bee’.

This was a reasonable idea, and Saffran and colleagues came up with a clever way to test it (Saffran, Aslin & Newport, 1996). They presented eight-month-old infants with a computer-synthesised stream of syllables like so:

…tupirobidakutupiropadotibidakugolabutupiropadotigolabubidaku…

The computer kept reading for three minutes with no pauses or variations in tempo and emphasis. The sequence was made from four nonsense words – tupiro, padoti, bidaku, golabu – randomly repeated. To see if the babies noticed this after the recording stopped, the researchers played them examples that were ‘words’, like ‘tupiro, golabu’ and other examples were ‘part-words’, with the end of one word and the beginning of another, such as ‘piropa, kugola’ and some ‘non-words’, which had the same syllables but in a completely new order ‘dapiku, tilado’. The results were clear. The babies, who we know love novelty, paid less attention to the familiar words than the other novel combinations.

This kind of statistical learning helps babies get their language learning underway. It allows them to segment speech into words before they even know what words are, building language from the ground up. You can do it too. If you listen to that computer language or a foreign language for long enough, you start to perceive the words. My colleague Bob French came up with a computer model that would learn in this way. Its full name is Truncated Recursive Autoassociative Chunk Extractor but we just call it TRACX (French, Addyman & Mareschal, 2011). The details of exactly what it does are not too important, except to say that it is a neural network that learns a lot like the brain.

We used TRACX to see if Saffran et al.’s idea scales up to real-world languages. First, we confirmed it could solve the simplified artificial grammar tasks researchers had used. Then we found a database of infant-directed speech, which included 9,800 phrases like ‘Look at the doggie’ or ‘Who’s on the telephone?’ We coded these phonetically with no spaces between the words and let TRACX ‘listen’ to the whole set. After several runs through, it was correctly dividing phonemes into words. We also found it had no problem with an artificial language problem we called ‘equal probabilities’, which confused other computer models. This led to a nice, testable prediction that if our model was realistic then babies ought to be able to learn the equal probabilities language too.

Barbara was one of the babies who came to the lab to test that prediction. Like Saffran et al., we set up a computer to read out our made-up language while the babies sat and listened. In our language there were six words – feego, feerou, neipau. neikoi, duhlu, duhtai – and the starting syllable was the same for three pairs of the words. This made a harder task than the original studies and this was why some other computer models failed. The babies had no problems: they could discriminate between words and part-words from our made-up language. We also were not concerned that Barbara was already multilingual. Our model was perfectly happy learning two different languages simultaneously. Classic studies of bilingual infants find they acquire words just as rapidly as monolinguals (Pearson, Fernández & Oller, 1993).

For babies, being bilingual seems be an advantage. There is a lot of literature that claims bilingual adults have an advantage in several important non-linguistic abilities such as executive control and working memory (Adesope et al., 2010), although some researchers dispute this, saying there has been a bias towards publishing only the positive results (de Bruin, Treccani, & Della Sala, 2015). But early in life the advantage seems clearer. Agnes Kovács and Jacques Mehler compared monolingual and bilingual seven-month-olds and found that the bilinguals showed an advantage in tests of executive control where babies had switch from one task to another (Kovács & Mehler, 2009). Natalie Brito and Rachel Barr found a memory advantage in six-month-old bilinguals (Brito & Barr, 2014). Curiously, in another study they found that bilinguals were better than monolinguals and trilinguals (Brito, Sebastián-Gallés, & Barr, 2015). However, the trilinguals do no worse than monolinguals and they have the undeniable advantage of speaking two more languages. If parents ask me if they should be exposing their baby to two, three, four or more languages I nod enthusiastically.

One interesting exception might be baby sign language. To my knowledge there are no studies of hearing children of a deaf parent learning sign language and a spoken language. But there is fierce debate about hearing parents of hearing children using a simplified baby sign language as a stepping stone to speech. Advocates argue that manual signs are easier than spoken words, so babies can express themselves sooner. Critics say this will delay babies speaking because they do not need to. The science itself is neutral. A meta-analysis of studies was inconclusive: there are no obvious benefits, but there were no downsides either (Fitzpatrick et al., 2014). So whenever parents ask me if they should use baby sign language I just shrug.

This excerpt is taken from Chapter 11 of my book, The Laughing Baby.