Before we can understand a word or a sentence, we must be able to perceive and put together the individual parts that make up each word. When we hear a spoken word, we are actually hearing a string of phonemes, which we then group together and separate into words through speech segmentation. Speech is not actually the production of distinct words, but a continuous stream of unbroken sound, which we then parse through speech segmentation. (This becomes evident when we listen to speech in a foreign language, which we cannot segment.) Phonemes are the smallest units of spoken language, the individual sounds that make up a word, such as [sh], [m], and [p]. Phonemes are not the same as syllables, because a single syllable often contains multiple phonemes. They are produced by manipulating air flow through the mouth and nose and vocalizing with parts of the throat.

There is some evidence that phoneme perception is an innate ability. By playing recorded phonemes to infants and measuring their rate of sucking on a pacifier, researchers found that even one-day-old infants can distinguish all the phonemes that humans can produce. Interestingly, humans seem to lose this ability over time; by the time they are one year old, children can only distinguish the phonemes used in their native language.

The phenomenon of categorical perception, used with reference to language, refers to humans' tendency to divide sounds into categories to create distinct phonemes where none really exist. For example, the phonemes [ba] and [pa] lie on a continuum; we can create a phoneme that sounds like 40 percent [ba] and 60 percent [pa], or any other such combination. When people are asked to listen to two phonemes and state whether they were the same or different, they find it much easier to distinguish the ba/pa combination phonemes if the phonemes straddle the 50 percent mark. A pair of phonemes in which one was 20 percent [pa] and the other 40 percent [pa] are difficult to discriminate, but a pair in which one is 40 percent [pa] and the other 60 percent [pa] is much easier, even though both pairs are separated by a difference of 20 percent. This suggests that there is a categorical boundary at 50 percent past which we perceive the phoneme as either "all-[pa]" or "all-[ba}," even though the range is acoustically a perfect continuum.

Phoneme perception can also be influenced by context. One study showed that if subjects hear a recorded word or sentence in which one phoneme has been deleted and replaced with static (white noise), they are unable to tell that anything is missing. They report hearing the entire word or sentence, including the missing phoneme. It seems that our auditory perception system knows which phoneme should have been there, and supplies it accordingly. For example, subjects might hear "The little child learned the al__abet quickly," but they will report hearing the [ph] phoneme even if they are told that it was deleted. This is an example of how context and prior knowledge (top- down effects) can influence perception.

But what about visual language? How do we perceive written words? Written language is composed of letters, rather than phonemes. These letters are, in turn, made up of lines and curves. These lines and curves are detected by feature detectors in the visual system of the brain. Each feature detector is set to detect a different, unique feature; for example, one may fire when it perceives a vertical line, and another may fire when a right angle is present. The output from these feature detectors is assembled into letters by a process similar to that of object recognition. For example, if the feature detectors for vertical line, horizontal line, and right angle fire, that output can be combined with information about those features' locations (from other areas of the visual system) to determine that we are looking at the letter "T." Once we know what letters are present and where they are located, we can combine them into words.

Like the phoneme detection system, the letter recognition system can be influenced by top-down effects of context. Look at the figure below.

The central letter in each word is exactly the same, yet we see one as an "H" and the other as an "A," based on their context. We know, based on our vocabulary, that the first word must be "THE" and not "TAE." Our recognition system uses this information to help correctly interpret the ambiguous letter as an H. The process is similar for "CAT"; we know that "CHT" is not a word, so the ambiguous letter must be interpreted and perceived as an A. Thus, our knowledge can direct our perceptions. Once we have assembled the letters into words, we can then tackle the problems of grammar.