History of Dictation: From Shorthand to Speech Recognition

Dictation began as a person speaking while someone else wrote. For most of history that someone was a scribe, and the speaker's words reached the page only through the scribe's hand. Three technologies have changed the arrangement since. Shorthand let trained writers keep up with speech. Recording machines let a speaker store a voice for a typist to transcribe later. Speech recognition, released as commercial software in 1997 ¹ and running on phone keyboards in 2019 ², lets the computer type what you say. Each change moved the writing closer to the person doing the thinking, and each one left some work behind.

Eight milestones cover the path from scribe to phone

The table gives the milestones this guide covers, each with the fact that supports it. The sections below explain each one and what it changed for the person writing.

Dictation began with people speaking to scribes

The practice is older than any writing machine. Plutarch's life of Julius Caesar says that during the Gallic campaigns Caesar "practised dictating letters on horseback and keeping two scribes at once busy, or, as Oppius says, even more" ⁹. Plutarch wrote long after Caesar's death and named his sources, so the detail is a reported account, not a record kept at the time. The point of the passage still holds: a commander with urgent letters could speak to a scribe and let the scribe carry the words.

To take dictation means to write down what someone says as they say it. It was the scribe's job, and later the shorthand writer's, in offices and courts. The speaker composes; the writer records. what "take dictation" means.

Dictation also shaped how books were written. Mark Twain had experimented with dictation as a method of composition as early as 1885 ¹⁰, and he began almost daily dictations to a stenographer in January 1906 ¹¹. The result was the bulk of his autobiography. For other writers who worked this way, see which famous authors dictated their books.

Shorthand let a trained secretary write as fast as an executive could speak

Shorthand is a system of symbols that lets a writer keep up with speech. It was how offices captured dictation as it was spoken. A stenographer wrote the symbols during the dictation and typed a clean copy afterward, the same two-step pattern that machines later copied. For the plain definition, see what shorthand dictation is. Isaac Pitman published his system, Stenographic Sound-Hand, in 1837 ¹², and it became the basis of Pitman shorthand. The pamphlet was printed in London at the press of Samuel Bagster ¹².

The speed a professional writer needs became a formal test. The National Court Reporters Association's skills test for the Registered Professional Reporter credential has a testimony leg at 225 words per minute ¹³. Each leg must be written at 95 percent accuracy to pass ¹⁴. Speed alone does not pass. A fast transcript with too many errors fails the leg, which is why the test pairs the two.

The chart below puts that testimony rate next to two measured rates for entering text. The three figures come from different tests, so the chart shows scale, not a head-to-head comparison.

The phone comparison is the one that matters for everyday writing. In a controlled study, English text entered by speech reached 153 words per minute on an iPhone, against 52 words per minute on the built-in keyboard ^{15 16}. The study counted the finished text, including corrections, which is why the gap is meaningful. The shorthand writer and the court reporter reached their speeds through training. The phone study measured short messages, not long documents.

Shorthand had a cost, though. The speaker still needed a writer, and the writer's notes still had to be typed up. That two-person arrangement became the standard office workflow. For how modern dictation differs from typing, see dictation versus typing.

Edison listed dictation as the first use of the phonograph in 1878

Thomas Edison's phonograph was new when he described its uses, and his list put dictation first. The Office Museum summarizes the list: in 1878 Edison "listed ten uses for the phonograph," and at the top was "letter writing and all kinds of dictation without the aid of a stenographer" ³. The list came from his article in the North American Review.

A Chicago Tribune report on that article, published in May 1878, described the same idea in business terms ¹⁷. Business people who dictated their communications to clerks could, with the phonograph, do without the clerk ¹⁷. The report's wording is a newspaper summary of Edison's proposal, not his own text, so treat the phrasing as the newspaper's.

Edison's office machines sold slowly at first. The Office Museum says neither Graphophones nor Edison Phonographs sold well for office use in the late 1880s or early 1890s ¹⁸. A slow start did not end the idea. The machines did reach offices, as the brand name below shows.

Edison's view of the market was specific. The IEEE History Center's entry on dictation machines says he believed the phonograph's primary use would be in business as a dictation machine ¹⁹. Dictation was the first use on his list. For what a dictation machine does today, see what a dictation machine is.

The Dictaphone name dates from about 1907 and became a generic word for dictation machines

The name was first used on Graphophone dictating machines. The Office Museum dates the trademark this way: "Around 1907 (in any case by 1908), Graphophone dictating machines began to use the Dictaphone trademark" ⁴.

The name then outlived the machines' dominance. Speech Technology magazine's history of the brand says Dictaphone dominated the market so thoroughly that "dictaphone" became the generic term for dictation machines ²⁰. When people in an office said "dictaphone," they meant any machine for recording dictation, whichever company made it. For a plain definition of the device, see what a Dictaphone is and the Dictaphone glossary entry.

Cylinders gave way to belts, tape and finally digital files

The storage medium changed several times. Each change made recordings easier to copy and store, and none of them changed the division of labor between speaker and typist. The table lists the milestones with their sources.

The Dictabelt is the change that matters most in this story. It cut a groove into plastic instead of wax, and the trade-magazine history says the company "phased them out in favor of 1947’s plastic carving Dictabelt" ⁵. Magnetic tape followed, and the same history says that magnetic tape and later hard drive machines became the dictation medium of choice ²³.

The cassette is the medium most readers will recognize. Philips presented its pocket recorder and compact cassette at the 1963 radio show in Berlin, and its museum describes a mini-cassette for dictation machines in 1967 ^{21 22}. The workflow stayed the same: the manager recorded, and a typist transcribed. A dictation machine records; it cannot type. For the device types that survive today, see what a dictation foot pedal is and the digital dictation entry.

In the office, the work was divided in two. The speaker dictated into a machine, and the typist loaded the recording onto a playback machine. The typist worked a foot pedal that controlled the speed of playback and let the typist reverse and repeat, and typed the letter on a typewriter ²⁴. The machine stored the voice, and the typist did the writing. Nothing in that arrangement let the machine understand the words.

The office history makes the same point. For most of the twentieth century, business people recorded voice letters, and a secretary typed them from the recordings ²⁵. A reader who wants the practical side of that workflow can start with how to dictate.

Speech recognition began as a lab problem and reached the desktop in 1997

Recording moved the voice. Speech recognition tried to move the typist's job onto a machine. The first milestone was AUDREY, built at Bell Laboratories. The Computer History Museum dates the start of its story to 1952, "at Bell Laboratories" ⁶. AUDREY could recognize the digits zero to nine "with 90% accuracy, but only if spoken by its inventor" ²⁶. Recognizing digits from one voice is a small problem, and the limit is the point: the machine could not handle a new speaker.

IBM's experimental Shoebox followed. IBM's archive says the device would "recognize 16 words spoken into its microphone" ⁷. Its vocabulary was still small, but it showed that a machine could respond to a set of spoken words.

Researchers in 1983 tested whether a computer could act as a typewriter that listens ²⁷. The authors wrote that speech recognition was "not yet advanced enough to provide people with a reliable listening typewriter" ²⁸. Their experiments used simulated listening typewriters, and they found that some versions, even on first use, "could be at least as good as traditional methods of handwriting and dictating" ²⁷. The recognizers in that study were simulated, so the result describes how people composed by voice, not how well software recognized speech.

The hard part is not the list of words. A speaker's voice changes with the speaker, the room, the microphone and the sentence, and a system that matches sounds to stored patterns has to allow for all of that. AUDREY dealt with the problem by being tuned to one voice, which is why its 1952 result was impressive and narrow ⁶. Systems that serve many speakers face a harder version of the same problem.

Dragon NaturallySpeaking, from Dragon Systems of Newton, was released in 1997 and cost $695 at launch, a price that later fell to $349 ¹. A 1997 Boston Herald article called Dragon Systems an industry leader ¹. For the first consumer tools and how they work, see the voice typing glossary entry and when computers first typed what people say.

Accuracy reached the human benchmark, and recognition moved onto phones

The next question was accuracy. Microsoft Research reported that its automatic system edged past the human benchmark on a conversational test set, with error rates of 5.8 percent on the Switchboard portion and 11.0 percent on CallHome ⁸. Professional transcribers made 5.9 percent errors on the Switchboard portion ²⁹. The test covers telephone conversations, not every kind of speech. Parity on that set does not mean the same result in a noisy room, with a heavy accent, or with technical vocabulary.

The next change was where the computation happened. In 2019, Google announced "an end-to-end, all-neural, on-device speech recognizer to power speech input in Gboard" ². That recognizer runs on the phone itself. Apple describes a similar arrangement for iPhone: Dictation there is processed on the device in many languages, and no internet connection is required ³⁰.

Availability still depends on the platform and the setup. Microsoft says Windows voice typing needs an internet connection, a working microphone, and the cursor in a text box ³¹. The history is not a straight line to offline dictation everywhere. If you need dictation without a network, check the platform's own settings before you assume it works. Yap, which we make, is an on-device dictation app for Mac and Windows. The offline dictation on Mac page covers the Mac side.

The modern dictation problem is correction. Speech recognition removed the typist, but the speaker now does the proofreading. A preliminary 2001 study of dictation with a speech recognizer found that users spent over 65 percent of their time correcting errors and under 35 percent dictating ³². That study used the recognizers of its time, so the figure is a warning about the workflow, not a measure of today's software. For the practical side, see dictation mistakes and how to fix them.

Each stage moved writing closer to the person doing the thinking

The history has a pattern. In each stage, one step between the thought and the text disappeared, or became something the speaker controlled. The scribe was the first step, and shorthand made the scribe faster rather than unnecessary. Recording machines let the speaker stop being present at the typing. Speech recognition removed the typist, which returned the writing to the speaker. Each change left the speaker with more responsibility for the final text, and the correction work moved toward the person who composed it.

Three conclusions follow for a reader today. First, the older tools are still in use: court reporters still write testimony in shorthand, as the NCRA's test shows. Second, the speaker still matters. AUDREY worked only for its inventor's voice, and the 2017 benchmark result ⁸ covers one test set, so how well a system handles your voice and your subject still decides the result. Third, correction is the part of dictation that most affects the outcome, so plan for it. For a comparison of speaking and typing speed, see speaking speed versus typing speed.

Several dates in this guide come from secondary histories, such as trade magazines, museum sites and encyclopedias, not from original documents. Where a date is approximate or only the year is given, the fact says so. For the broader picture of what dictation is, start with what dictation is and the dictation overview.