speech recognition (ASR)

in English: speech recognition / speech-to-text

basicwidespread

Speech recognition is a technology that automatically converts spoken human language into written text, allowing the machine to "transcribe" what you say.

In plain terms

Imagine a tireless stenographer who types out everything they hear, whether dictation, a phone call, a lecture, or a video. Speech recognition does the same as software: sound goes in, text comes out.

It serves to make spoken words searchable, editable, and shareable. This becomes the automatic subtitles for videos, the transcript of meetings, and the voice-dictated message on the phone. It is often the first step even when you are conversing with an AI assistant via voice.

It is important that speech recognition only transcribes, it does not interpret. What the transcribed sentence means and what to do with it is decided by another system, such as a language model or an AI agent. Some newer, so-called multimodal models can process sound directly without transcription, but this is already more than just speech recognition.

How it works

The microphone converts sound into a digital signal, and the software breaks it into short slices of a few hundredths of a second.
For every slice, it calculates how strongly high and low vibrations resonate within it. From this, a visual fingerprint of the sound is created, showing how the sound changes over time (called a spectrogram).
A neural network trained on a vast number of hours of transcribed audio recordings estimates which sounds, syllables, and words the audio samples correspond to.
The model also considers the context: among similarly sounding options, it chooses the one most likely in that specific sentence.
Modern systems add punctuation and capitalization to text, often include timestamps, and some solutions also indicate who is speaking.

An everyday example

You are rushing for the bus with no hand free to type, so you press the microphone button on your phone and dictate: "I will be ten minutes late, start without me." The phone transcribes the sentence, and you send it with one tap.

Who is affected

everyonePhone dictation, voice-based search, and automatic video captions all rely on this, so you use it almost daily.

employeeAutomatic transcripts and minutes from online meetings save much manual note-taking. → Work & HR

patient, doctorA doctor can dictate findings, but a misheard drug name or dosage can cause serious errors, so the transcript must always be checked. → Health

parent, student, teacherLecture transcripts and live captioning support learning and make the class accessible to people with hearing impairments. → Education & Family

A common misconception

MythIf the machine transcribed exactly what I said, it understood it.

In realitySpeech recognition converts speech into text. It uses context only to recognize words well, not to understand the meaning of sentences. Understanding and responding are the job of another system, and the transcript can contain errors even if it reads smoothly.

Limits and risks

Accuracy noticeably declines in noise, with poor microphones, when speakers interrupt each other, with strong dialects, or with technical jargon.
Names, numbers, abbreviations and foreign words are the most common sources of error, so always read through a text (a contract, a medical report, a quotation) carefully.
Some models may insert text in quiet or noisy sections that was never spoken.
Your voice is personal data: check whether the recording stays on your device or goes to a company server, and record others' speech only with their knowledge.

Similar, but not the same

speech synthesis – Speech synthesis works in the opposite direction: it generates machine voice from written text, while speech recognition creates text from voice.
speaker recognition – Speaker recognition determines who is speaking, while speech recognition determines what is being said.
natural language understanding – Language understanding interprets the meaning and intent of already written text, while speech recognition only reaches the transcription stage.

Related concepts

Broader

artificial intelligence, Natural language processing, machine learning

Narrower

dictation, automatic captioning, voice control, speaker diarization

On the edge

speech synthesis, voice assistant, voice cloning, multimodal model, machine translation

A brief history

The first known speech recognizer was the machine named Audrey, introduced by Bell Labs in 1952, which recognized spoken digits from a single speaker. OpenAI released the open-source Whisper model in 2022, which made usable accuracy widely available in many languages, including Hungarian.

Where things stand

At the end of 2026, speech recognition is a basic feature of phones, video-calling software, and AI assistants, and it works well in Hungarian too. Development is moving toward real-time voice conversation, multimodal models that process audio directly, and models that run on the device without an internet connection.

Status: 2026 November

In the news

Appears in 3 Agent Age stories; the curve shows weekly occurrence in the news magazine.

Multilingual medical speech recognition: Oak and other AI toolsOct 13, 2026 · Health
Hungarian Bar Association issues position on lawyers' use of AI speech-to-text toolsAug 10, 2026 · Jog
CLIVORA project develops Hungarian-language AI medical speech recognition and report generationAug 3, 2026 · Health

Also known as

speech recognition · Automatic speech recognition · speech-to-text conversion · speech recognition · speech recognition · speech-to-text · STT · ASR

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.