IELTS Academic Reading Practice Test 111

In the summer of 1877 Thomas Edison was working at his laboratory in Menlo Park, New Jersey, on a machine that would record the dots and dashes of the telegraph on a strip of paper. He noticed that when the paper was pulled through the device quickly, the marks produced a faint humming noise, and it occurred to him that a needle following a groove might reproduce sound as well as record it. By the end of November he had sketched a design, and a few weeks later his machinist, John Kruesi, built the first working machine. Edison later recalled that the first words spoken into it were the nursery rhyme 'Mary Had a Little Lamb'.

The instrument was simple. A brass cylinder about ten centimetres long was mounted on a shaft turned by a handle, and its surface carried a continuous spiral groove. Over the cylinder projected a small metal point, or stylus, attached to a thin diaphragm held in a frame. When someone spoke into a mouthpiece, the sound waves pressed against the diaphragm and made it vibrate, and the stylus, travelling along the rotating cylinder, pressed a series of indentations into a sheet of tinfoil wrapped around it. To hear the recording, the operator turned the cylinder back and the same point ran through the indentations, setting the diaphragm in motion once more.

Edison filed a patent for the invention in December 1877, and it was granted in February 1878. He was convinced that the phonograph would be used above all in offices: its most valuable service, he wrote in an article published in 1878, would be to record letters dictated by a businessman and typed later by a clerk. His list of possible uses also included talking books for people who were blind, the teaching of elocution, the preservation of family voices and, near the bottom, music. Edison, who was somewhat deaf, had little interest in music and did not expect that the public would buy the machine chiefly to listen to songs.

The tinfoil phonograph was a scientific marvel and a commercial failure. The foil tore easily, and after a handful of playings the indentations blurred and the sound became indistinct. Each cylinder was a single original that could not be copied, so a recording survived only as long as its foil did. Edison himself soon turned his attention to other projects, and for almost a decade the phonograph remained a curiosity demonstrated in lecture halls rather than a product anyone could buy.

Other inventors saw the potential. Working in Washington, Alexander Graham Bell joined with his cousin Chichester Bell and the instrument-maker Charles Sumner Tainter to improve the machine. Their version, called the graphophone, replaced the tinfoil with a coating of wax on a cardboard or metal cylinder. Wax was soft enough for a stylus to cut a clean groove and firm enough to survive repeated playing, and the surface could be shaved flat with a knife so that a cylinder might be used again. The three men patented their system in 1886 and formed a company to develop it.

Meanwhile a German-born inventor living in the United States took a different route. Emile Berliner had already devised an improved microphone and sold the patent to Bell's company. In 1887 he patented a machine he called the gramophone, which recorded on a flat disc of zinc coated with a thin layer of wax. Whereas the phonograph and the graphophone pressed vertical dents into a cylinder, Berliner's stylus moved from side to side, cutting a groove of varying width. The advantage of a disc was not superior sound but copying: a single recording could be etched and pressed to produce thousands of identical copies, a process far cheaper than recording each cylinder separately.

For more than a decade, however, the cylinder held its own, partly because early discs were small and rough-sounding. Public listening became the fashion. From the late 1880s entrepreneurs installed coin-operated phonographs in saloons, arcades and railway stations; for a nickel a customer heard a short performance through rubber tubes, and several listeners could share the ear-pieces at once. In the first years of the twentieth century the disc machine, with its cheaper records, displaced these devices in the home, and companies such as the Victor Talking Machine Company, founded in 1901, built cabinets in which the horn was concealed inside the furniture.

The next transformation came from electricity. Until the mid-1920s all recording was acoustic: performers played into a horn, and the energy of their sound alone drove the cutting needle. From 1925 the record companies began to adopt microphones and valve amplifiers developed by the telephone industry. Electrical recording captured a far wider range of frequencies, allowed a singer to stand some distance from the instrument, and made it possible to balance the parts of an orchestra.

Edison resisted the trend, insisting that cylinders gave a truer reproduction, but his company stopped making them in 1929. Discs then reigned unchallenged until the long-playing record arrived in 1948, offering about twenty minutes of sound on each side instead of the four or five typical of the old shellac discs, and pressed in the tougher material polyvinyl chloride. The phonograph's basic insight, that a groove can hold a vibration which a needle can later restore to sound, had by then been carried into tape recording, and much later into the digital files that now dominate listening.

Every time a passport is read at a border, a postal sorting machine scans an address, or a scanner turns a page of a printed book into searchable text, optical character recognition (OCR) is at work. The task sounds simple: look at a picture of writing and say what it says. In practice it draws on several branches of computer science, because a machine has no innate notion of what a letter is. To a computer, a scanned page is a grid of picture elements, or pixels — typically 300 to the inch — each carrying a number that records how dark it is. A modest scanner produces about eight million such numbers for a single page of A4.

The first stage is to clean the image. Scanners rarely produce perfect pictures: pages are fed in slightly crooked, paper may be stained or yellowed, and photographs bring shadows and uneven light. Programs therefore convert the image to black and white by choosing a threshold, remove specks of noise, and measure the angle at which the lines of text run so that they can be rotated back to the horizontal, a step known as deskewing. Good preprocessing is unglamorous but decisive, since a badly skewed page will defeat the cleverest recogniser.

Next comes segmentation, the division of the page into its parts. The software must decide which regions contain text and which contain pictures, and must then separate columns, lines, words and finally individual characters, or, in joined-up writing, strokes. On a modern page with tables, captions and sidebars this layout analysis is often harder than recognising the letters themselves. Forms and invoices, where a value may sit in a box to the left of its label, cause particular difficulty. Historical documents present worse problems, since printers of the past mixed typefaces freely, ran words together and used abbreviations that no modern dictionary lists.

Recognition itself has been approached in two broad ways. The older method, template matching, compares each character image with stored examples and chooses the closest fit. It works well with a single typeface printed cleanly, which is why the first commercial machines were trained on a handful of typewriter faces and failed when asked to read anything else. The alternative, feature extraction, describes each shape by measurable properties — the number of holes enclosed by a loop, the position of line endings, the direction of curves — and then classifies it. Feature methods cope better with unfamiliar fonts and with the distortions introduced by scanning.

Because no recogniser is perfect, most systems finish with post-processing. A dictionary check can correct an isolated error: the sequence 'teh' is not a word, and 'the' is the obvious candidate. Statistical language models go further, estimating which sequences of words are plausible in the language as a whole, so that a doubtful character can be resolved from its neighbours. This is why OCR engines are usually supplied with a list of the languages they may be asked to read, and why a document in an unexpected language produces strange results.

The idea is older than the computer. In the early 1910s the physicist Edmund Fournier d'Albe built the optophone, a device that scanned printed letters with a beam of light and turned the reflected light into tones of different pitch, so that a blind reader might learn to interpret the sound. Reading was extremely slow and few people mastered it, but the principle of converting print into a signal a machine could handle was established.

Practical reading machines appeared in the 1950s, when computers became powerful enough to analyse a page. Early devices were enormous, could handle only a few specially designed typefaces, and were sold mainly for routine clerical work: reading electricity meters, sorting mail, processing insurance forms. In 1976 Ray Kurzweil, an American inventor, launched a reading machine for blind people that combined a camera, OCR software and a speech synthesiser, and it was taken up by users, among them the musician Stevie Wonder.

OCR became cheap and ubiquitous over the following decades. Hewlett-Packard began developing the Tesseract engine in the mid-1980s and released it as free software in 2005, after which it was maintained and improved by Google and a community of volunteers. Since about 2015 the field has been reshaped by deep learning: neural networks trained on millions of examples learn the features for themselves instead of being told which properties to measure, and recurrent networks can read a whole line of text at once rather than one character at a time. Accuracy on clean modern print is now high, and error rates of around one per cent are common. Phone apps perform the same steps in a fraction of a second on images captured by hand.

Yet difficulties remain. Handwriting is far harder, because no two writers form a letter in the same way; so is text photographed in the street, with its odd angles, shadows and busy backgrounds. Torn or faded documents produce errors that can silently corrupt an archive, so libraries increasingly insist on a second pass in which a human compares the text with the image. Degraded print from the nineteenth century, scanned at low resolution, is a further challenge. Meanwhile, work continues on scripts and languages for which little digital material exists, since there is not enough data to train the models that serve English so well.

A decade ago the smart speaker was a novelty. Amazon's Echo, launched in the United States in November 2014, was followed by Google Home in 2016 and Apple's HomePod in 2018, and by the early 2020s the devices stood in a substantial minority of homes in Britain and North America. The object itself is unremarkable: a short cylinder or a small dome containing a handful of microphones, a loudspeaker, a light and a connection to the internet. It waits for a wake word, records what follows, sends the audio to a distant server and speaks the answer that comes back. Children, who can operate one long before they can read, took to it immediately.

Anyone who has watched a five-year-old interrogate a voice assistant will recognise the pattern. The questions come in a flood: why is the sky blue, what do spiders eat, how many bones has a giraffe. For a parent cooking dinner, the appeal is obvious. The device never tires of answering, never asks for quiet, and can play the same song for the fortieth time without irritation. For a child who cannot yet read, or who is learning English as a second language, the ability to summon information by voice is genuinely new.

Yet the word 'conversation' is doing a great deal of work in the marketing of these devices. What passes between a child and a smart speaker is not a conversation but a series of transactions: one question, one answer, no follow-up, no shared attention to a third thing, no misunderstanding repaired. The features that make talk with an adult so valuable to a young child — the pauses, the gestures, the willingness to be interrupted — have no equivalent in the machine. A device that answers every question may, in a small way, reduce the incentive to ask a person.

There is a practical difficulty too. Speech recognition systems have been trained overwhelmingly on adult voices, and children's voices are different: higher in pitch, less precise in pronunciation, and delivered at unpredictable speed. Researchers who study the problem report that recognition rates for young children are noticeably worse than for adults, and that the gap is widest for the youngest speakers. It has been suggested that some children learn to shout, or to flatten their vowels, in order to be understood, which is a curious outcome for a technology sold as effortless.

Privacy is the issue that ought to worry parents most. A device that listens for a wake word is, by design, a microphone in a bedroom or a kitchen, and the recordings it makes are usually processed in distant data centres. In 2019 it emerged that contractors working for Apple and Amazon had been listening to short excerpts of customers' recordings in order to improve the systems. Apple suspended the practice, and both companies later changed their settings so that users could opt out. In the United States, legislation passed in 1998 requires companies to obtain parental consent before collecting personal information from children under thirteen, and Britain's Age Appropriate Design Code, in force since 2021, sets stricter defaults for services likely to be used by children.

Commercial pressures add another layer. The speaker is cheap because it is a gateway: to music subscriptions, to audio books, to shopping, to games that require a monthly payment. Devices sold for children come with parental controls that block purchasing and filter content, which is welcome, but campaigners have complained that some applications aimed at young listeners carry advertisements, and that the line between a story and a sales pitch is not always clear to a six-year-old.

The benefits should not be dismissed. For a child with a visual impairment, a voice interface opens a library. For a child with dyslexia, hearing a word spoken aloud while seeing it written can help. Teachers of children with autism report that a device which gives the same answer in the same tone, and can be asked to set a timer or repeat a routine, is sometimes calming rather than threatening. Yet the evidence that such devices produce measurable educational gains is thin, and most claims rest on anecdotes rather than on controlled studies.

My own view is that the machine is not the problem. The risks that matter are created by design decisions — how long audio is kept, who may listen to it, what is done with the data, whether children are marketed to — and by adult habits, since a speaker that keeps a child occupied is a speaker that keeps a child out of the way. A household in which the device is treated as a companion, left in a child's bedroom and never discussed, is a different proposition from one in which it is treated as a radio with a memory.

Nor do I have much patience with the call for bans. Families that remove the speaker do not thereby create conversation; the television and the tablet are still there. What is needed is regulation of the defaults, so that audio is processed on the device where possible, deleted quickly when it is not, and never used for advertising to children, together with a good deal of ordinary adult effort. The question is not whether children should talk to machines, but what the adults in the room are doing while they do.