Dragon Dictate

What Dragon Dictate was

Dragon Dictate was the first speech-recognition product made for consumers. Dragon Systems released it in 1990, and it sold for $9,000^[1]. The idea behind it is easy to state and hard to build: a person speaks, and the computer writes down what was said, so that a document is produced by voice instead of by keyboard.

The discontinuous era it came out of

Earlier speech-recognition programs all used discontinuous dictation, which meant the user had to pause after every word^[1]. Speech was handled one word at a time, and the machine never had to work out where one word ended and the next began inside a stream of sound. The price of that simplification was speed: dictation became a slow, halting affair of speak, pause, speak, pause.

Research on machine speech recognition had been going on in laboratories well before this. In the 1980s IBM built a voice-activated typewriter called Tangora, with a vocabulary of 20,000 words^[1]. IBM's approach was based on hidden Markov models, and that brought statistics into digital signal processing^[1]. Rather than hand-writing rules for how each sound should be identified, a statistical model treats the incoming audio as evidence and estimates which sequence of sounds and words most probably produced it.

Continuous speech

The change that mattered most to people actually writing documents came in 1997, when Dragon NaturallySpeaking appeared as the first continuous speech recognition product^[1]. Continuous recognition means the speaker no longer stops after each word; the software itself has to find the boundaries between words in ordinary running speech. NaturallySpeaking could recognise 100 words per minute^[1]. Peter Mahoney, senior vice-president of Nuance Communications, described Dragon as the pioneer of continuous speech recognition and said it made practical speech recognition usable for creating documents^[1].

Where it sits

Dragon Dictate marks the point at which speech recognition stopped being only a laboratory subject and became something a customer could buy^[1]. At $9,000, that first version was priced far beyond an ordinary purchase, so it was a tool for people and organisations that could justify the cost rather than a mass-market product. What it established was the direction: a system that turns dictation into text you can edit and print. The problems left open in 1990 — the pause after every word, the limited pace of input — were exactly the ones that continuous recognition addressed later on^[1].

Seen from that angle, the product is less interesting as a finished piece of technology than as the opening of a path. The vocabulary of Tangora, the statistics of hidden Markov models and the consumer packaging of Dragon Dictate all point at the same goal: a machine that takes spoken language and gives back a document^[1].