Speech recognition

In short

Speech recognition is technology that turns spoken words into written text. It powers dictation, live captions, voice assistants and meeting transcripts, and modern versions use AI models trained on huge amounts of recorded speech.

Also called: Automatic speech recognition, ASR, Speech-to-text, STT, Voice recognition, Transcription
Topic: How AI models work
Level: Beginner
First described: Davis, Biddulph and Balashek of Bell Telephone Laboratories described an automatic recogniser for spoken digits in the Journal of the Acoustical Society of America in November 1952¹.

A speech recognition system takes audio and outputs a transcript. Early systems were narrow. A 1952 machine built at Bell Telephone Laboratories recognised the ten spoken digits from one person, with 97 to 99 percent accuracy, and had to be adjusted for each new speaker¹.

Modern systems are neural networks trained on vast amounts of recorded speech paired with transcripts. OpenAI's Whisper, published in 2022, was trained on 680,000 hours of audio from the internet in many languages². Its authors report that it handles standard tests well without extra fine-tuning and approaches human accuracy and robustness².

Through OpenAI's API, Whisper supports 98 languages, though OpenAI says accuracy varies by language³. Newer transcription models on the same service can also label who is speaking³.

Speech recognition only works out which words were said. It does not by itself understand them or check them. Many products now pass the transcript to a language model to summarise it, which adds a second place where errors can creep in.

In the NHS, "ambient scribe" tools combine speech recognition with generative AI to turn a consultation into structured notes and letters, and NHS England has published guidance on using them⁴.

Example

A GP uses an ambient scribe during an appointment. It listens to the conversation, transcribes it and drafts the clinical note for the doctor to check, the kind of product NHS England's 2025 guidance covers⁴.

Why it matters

It is how you talk to your phone, get captions on a video call and get meeting notes without typing. Its mistakes end up in written records, so transcripts that matter still need a human check.

Questions people ask

Is speech recognition the same as voice recognition?

People often use the terms interchangeably. Working out who is speaking is a different job from working out what was said, and some transcription models now do both, labelling which speaker said each part³.

How accurate is AI speech recognition?

On clear audio, very good. Whisper's authors report it approaches human accuracy and robustness². But results depend on the recording and the language: OpenAI notes Whisper's accuracy varies by language³. Check names, numbers and technical terms in any transcript that matters.

Does the NHS use AI to transcribe appointments?

Ambient scribe tools that turn spoken consultations into notes and letters are being introduced in health and care settings. NHS England first published guidance on their use in April 2025 and has since updated it⁴.

Related terms

Tools that use it

Sources

Back to the glossary