Who invented speech recognition?

No single person invented speech recognition. It was built in steps by research teams at Bell Laboratories and IBM, and later by university and company groups. Each step solved a different problem: first recognizing digits, then a small set of words, then continuous speech. Bell Laboratories built an early digit recognizer in 1952.

Bell Laboratories built a digit recognizer in 1952 that heard only one voice

Juang and Rabiner credit Davis, Biddulph and Balashek of Bell Laboratories with a 1952 system for isolated digit recognition for a single speaker ¹. The Computer History Museum describes AUDREY, a digit recognizer from the same period, which recognized the digits zero to nine with 90 percent accuracy, but only when its inventor spoke ².

That speaker limit is why "inventor" is the wrong word for these early systems: each one was built for a single voice.

IBM's Shoebox was credited with 16 spoken words in 1961

IBM says engineer William C. Dersch introduced its experimental Shoebox machine at a press conference in 1961 ³. Shoebox is usually credited with 16 spoken words ⁴. The step mattered because it went beyond digits to a small set of words.

IBM's effort, led by Fred Jelinek, was one of two schools of thought

Juang and Rabiner describe the 1970s as a time when IBM and AT&T Bell Laboratories were two different schools of thought on commercial speech recognition, and they say IBM's effort was led by Fred Jelinek ⁵. Our glossary entry on the hidden Markov model explains one of the statistical models of this period.

Deep neural networks in 2012 markedly reduced error on a telephone benchmark

A 2012 review by a group of authors reported that a deep neural network reduced the word error rate of a baseline system from 27.4 to 18.5 percent on a Switchboard telephone task ⁶. The number summarizes several systems, so treat it as a representative gain, not a single controlled test.

In 2006, Graves and colleagues presented a method that trains recurrent networks to label unsegmented sequences directly. Earlier networks had needed pre-segmented training data ⁷. The method is known as connectionist temporal classification.

The answer depends on which invention you mean

If you mean the first digit recognizer, the answer is Bell Laboratories in 1952. If you mean the first general-purpose, large-vocabulary continuous product, the answer is Dragon Systems and its NaturallySpeaking product, unveiled in 1997 ⁸. If you mean the open model you can download and run yourself, it is OpenAI's Whisper, described in a 2022 paper that lists six authors, among them Greg Brockman and Ilya Sutskever ⁹.

Because the field has many firsts, a single name is always a simplification. Say which step you mean, and cite the paper or the date that supports it.

What to do if you need a citation

Use the primary sources. Juang and Rabiner's 2005 history traces the programs above ⁵. Our history guide lists each milestone with its own source. If you are writing for school or work, name the team and the year rather than one person, and link to the paper.

For the papers behind these steps, see our question on important speech recognition papers. To learn the theory yourself, see our question on how to learn speech recognition. For how today's systems are trained, see how speech recognition models are trained. The full timeline is in our overview of speech recognition.