Speech recognition is the general term for technology that identifies spoken words and converts them into text or commands. It is used interchangeably with automatic speech recognition and speech-to-text. The phrase covers both full transcription and narrower tasks like recognizing a spoken command or a wake word, which is the main way it differs in scope from the strictly transcription-focused labels.
A brief history of speech recognition
Speech recognition research goes back to the 1950s. Bell Labs built "Audrey" in 1952 to recognize spoken digits, and IBM demonstrated the 16-word "Shoebox" in 1962. A 1970s DARPA program produced Carnegie Mellon's Harpy, which handled a vocabulary of around 1,000 words.
Statistical methods took over in the 1980s and 1990s, when hidden Markov models made larger vocabularies practical and products like Dragon Dictate reached consumers. Deep neural networks and then end-to-end models drove the sharp accuracy gains of the 2010s, and assistants like Siri put speech recognition in front of a mass audience.
Transcription versus command recognition
The two ends of speech recognition look different in practice. Open-vocabulary transcription, formally called large-vocabulary continuous speech recognition, turns free-form speech into text and is what a speech-to-text API does.
Command-and-control recognition works from a small, fixed grammar, matching speech against a limited set of allowed phrases, which is how older phone menus and simple voice controls worked. Keyword spotting and wake-word detection are narrower still, listening only for a specific trigger like a device name.
How speech recognition works
Across both command and open-vocabulary systems, the speech recognition pipeline is similar. The system extracts acoustic features from the audio, then a model predicts the words.
Command systems constrain that prediction to a known grammar, which makes them fast and accurate within their vocabulary but useless outside it. Open transcription systems use a full language model, or an end-to-end neural network, so they handle any sentence at the cost of more computation.
Accuracy on open transcription is reported as word error rate.
The speech recognition naming overlap
For general use, "speech recognition," "ASR," and "speech-to-text" mean the same thing, and vendors mix the terms freely. The one label to read carefully is voice recognition.
In its strict sense, voice recognition means identifying who is speaking, a biometric task, not what they said. When a product says "voice recognition" it usually just means transcription, so confirm which sense is meant, because the two solve entirely different problems.
Frequently asked questions
What is speech recognition? Speech recognition is technology that enables computers to identify spoken language and convert it into text or commands. It is also known as automatic speech recognition or speech-to-text.
How does speech recognition work? The system extracts numerical features from the audio, then a model predicts the corresponding words. Open transcription uses a large language model or end-to-end network; command systems match speech against a fixed grammar.
Is speech recognition the same as voice recognition? Not strictly. Speech recognition identifies the words that were said, while voice recognition, in its precise meaning, identifies the person speaking. Many products use "voice recognition" loosely to mean transcription, so check which sense applies.
Is speech recognition AI? Modern speech recognition is built on machine learning and neural networks, so it is a form of AI. Earlier systems relied on statistical models like hidden Markov models, which are also machine learning but predate the deep learning era.