Learn AI /Generative Media /Speech-to-text
Speech-to-text
- Starter
- Generative Media
- Communications teams
- Programme & data teams
Also called: automatic speech recognition, ASR, transcription.
Back to the milestone you came from
In five seconds
What it is
Tools in this group turn spoken audio into written text, while related labels differ slightly in scope, method and how much human checking is involved.
In plain words
It turns spoken audio into written words. The system makes a text draft from speech in the audio, and a person may need to correct mistakes.
An everyday example
Imagine a meeting in Amman where one person speaks and a note-taker types almost instantly on a laptop. The note-taker listens for word sounds, uses the sentence around them to guess unclear parts, and adds commas so the notes read smoothly. If two people talk at once or a name is unusual, the note-taker may make mistakes and someone checks the notes afterwards. The final notes are the transcript, even if the first draft was produced very quickly.
A human note-taker uses general world knowledge, intentions and meeting context in richer ways than a speech recognition system usually does. Real ASR works by statistical pattern matching on audio and text data, not by listening like a person. Also, some systems output time stamps, confidence scores or speaker labels, which the note-taking story does not show well.
Remember thisSpeech-to-text gives you a fast text draft from audio, but a transcript is only trustworthy after checking the parts that matter.
Understand it
Understand it
Why it matters
Speech-to-text can make spoken information easier to search, review, subtitle and reuse. It can help people who prefer reading, support note-taking, and speed up routine documentation. But it can also introduce errors that change names, numbers or meaning. For non-specialists, the key decisions are practical: whether the tool is accurate enough for the purpose, whether a human must check it, whether the audio can legally and safely be processed, and whether users might wrongly treat the transcript as exact. It is useful, but it is not the same as a verified record.
From spoken audio to checked transcript
Show the difference between raw audio, automated recognition, and the final transcript after review.
01
1. Audio in
A microphone or recording provides spoken language as the input.
Draw a box with a waveform icon and the label 'Live speech or recording'. Add small examples: 'meeting', 'interview', 'voice note'.
02
2. Speech recognised
The system analyses the sound and predicts which words were spoken.
Draw a processing box with simple waveform-to-word symbols. Add a note below: 'accent, noise, overlap and names affect accuracy'.
03
3. Draft transcript
The output is draft text, sometimes with punctuation, time stamps or speaker labels.
Draw a document box with three sample lines of text. Add tags reading 'punctuation', 'timestamps', 'speaker labels'.
04
4. Review and final text
A person may correct mistakes before the text is used or shared.
Draw a final document box with a checkmark. Add an optional side icon of a person editing, connected to the draft transcript box.
A left-to-right flow diagram with four boxes connected by arrows. It starts with audio input, then automated recognition, then draft text, then optional human review leading to final transcript. Small notes under the middle boxes show where errors can enter and where corrections happen.
On hover or tap, each box reveals one typical source of error or one quality improvement step; with keyboard focus, the same text appears below the diagram.
Describe this without the diagram
This diagram is a four-step flow from speech to finished text. Step 1 is audio input such as a meeting recording or live microphone. Step 2 is automated speech recognition, where the system analyses sound patterns and predicts words; accents, background noise, overlapping speech and unusual names can reduce accuracy here. Step 3 is a draft transcript, which may include punctuation, time stamps or speaker labels. Step 4 is optional human review, where a person corrects errors before the transcript is used as the final text. The main lesson is that automated recognition and a finished transcript are related but not identical.
With reduced motion: Show the same four boxes and arrows as a static diagram, with short notes under each stage and a visible side note that human review is optional but often important.
Speech-to-text is the broad idea of converting spoken words in audio into written text. The input is usually a live microphone feed or a recording. The output is text, often with time stamps and sometimes speaker labels.
Automatic speech recognition, or ASR, usually refers to the automated part of that process. A system analyses patterns in the sound signal, breaks speech into small pieces, and predicts the most likely words. In simple products, people often use “speech-to-text” and “ASR” as near-synonyms.
Transcription is slightly broader. It can mean the finished written record of speech, and it can be created by a person, by software, or by both together. So all ASR can support transcription, but not all transcription is fully automatic.
What happens between input and output depends on the system. It may detect where speech starts and stops, estimate which sounds are being spoken, use a language model to choose likely word sequences, and add punctuation. Accents, background noise, overlapping speakers and uncommon names can reduce accuracy.
This term is often confused with machine translation and audio generation. Speech-to-text writes down what was said in the same language unless translation is added as a separate step. It also does not create a synthetic voice; that is a different task.
How “A fast note-taker in a meeting” maps onto the real thing
| In the story | In the AI system |
|---|---|
| The speaker in the meeting | The incoming audio signal |
| The note-taker hearing sounds and typing words | Automatic speech recognition converting audio patterns into text |
| Using the sentence to guess an unclear word | A language model helping choose the most likely word sequence |
| A person reviewing the notes afterwards | Human correction in transcription workflow |
A worked example
A reporting officer in a regional bureau transcribes a short extract from a recorded briefing for a background note due for clearance.
Starting from: Audio: “The revised paragraph is ready for clearance. Please place the signed note on the file before close of business.”
- The system identifies the portions of the waveform that contain speech.
- It analyses acoustic features and predicts the most likely sequence of words or tokens.
- It adds punctuation and time alignment. If enabled, speaker diarisation estimates who spoke when.
- The draft renders “on the file” as “on the five”. The reporting officer checks the audio and corrects the phrase.
Result: Checked transcript: “The revised paragraph is ready for clearance. Please place the signed note on the file before close of business.”
In a United Nations workplace
An evaluation officer supporting a UN Country Team working group uses an approved speech-to-text tool to draft a transcript of the recorded discussion. The officer checks it against the audio while finalising the situation report before the regional bulletin closes.
The draft helps the officer locate interventions quickly, but overlapping speech, unfamiliar names and poor audio can alter wording or assign a comment to the wrong speaker. Names, figures, quotations and attribution must be checked against the recording.
Before you try thisRecordings containing personal, protection-sensitive, security-sensitive or other confidential material must not be uploaded to a tool that is not approved for that material. Doing so could expose individuals and breach handling requirements. Access to the audio and transcript must also be restricted appropriately.
Where you may hear it
- Live captions during a staff town hall
- Preparing the record after a virtual inter-agency working group
- Reviewing a recorded press briefing before quotations are cleared
- Testing accessibility arrangements before a side event
- A vendor demonstration during a technology governance board meeting
Avoid confusion
Avoid confusion
Often confused with
Audio and speech generation
Tools in this area create new speech, music or other sounds from text, audio examples or other controls, including voices that imitate a particular person.
Machine translation
Machine translation uses computer systems to convert text from one language into another using learned patterns, statistical methods or hand-built rules.
Text generation
A system produces new written words from a prompt, patterning its output on examples seen during training.
Common misconceptions
Common misconception: Speech-to-text and transcription mean exactly the same thing.
Correction: They overlap, but they are not identical. Speech-to-text or ASR usually points to the automated conversion step. Transcription can mean the broader process or the final written record, which may include human correction.
Common misconception: If the transcript looks fluent, it must be accurate.
Correction: Systems often produce smooth text even when specific words are wrong. Names, figures, places and technical terms are common failure points, so readability is not proof of correctness.
Common misconception: Speech-to-text automatically translates speech into another language.
Correction: Usually it does not. Standard speech-to-text writes down speech in the same language. Translation, if offered, is typically a separate step or a separate system layered on top.
Quick check
Which statement best explains how these terms relate?
Option 1 is the best answer. In everyday use, speech-to-text and ASR are often used almost interchangeably, but ASR usually names the automated recognition mechanism. Transcription is broader and can mean the final written record, whether produced by software, a person, or both.
Go deeper
Go deeper
The simple version
It turns spoken audio into written words. The system makes a text draft from speech in the audio, and a person may need to correct mistakes.
The technical version
In technical use, ASR refers to automated conversion of an audio signal into a sequence of words or tokens. A system processes acoustic features from the waveform, estimates probable speech units, and combines that with a language model or other sequence modelling method to predict the most likely text. Many products also add punctuation, time alignment and speaker diarisation, meaning an estimate of who spoke when. “Speech-to-text” is the broader user-facing term for this function. “Transcription” is broader still: it can refer to the output document or the whole workflow, whether generated automatically, manually, or through human correction of an ASR draft. This should be distinguished from speech translation, which changes language, and from text-to-speech or other audio generation, which create audio rather than extract text from it.
Reviewed external resources
Primary sourceRecommendation on the Ethics of Artificial Intelligence (opens in a new tab)UNESCO
This UNESCO recommendation sets ethical principles for AI use and suits readers who need a governance and rights-based frame rather than a product tutorial.
Intermediatelink checked 2026-07-31
Primary sourceContent Credentials (opens in a new tab)Content Authenticity Initiative
This Content Credentials resource helps readers understand provenance tools for digital media and suits those interested in authenticity questions around AI-generated and edited content.
Starterlink checked 2026-07-31
EncyclopaediaSpeech-to-text reporter (opens in a new tab)Wikipedia
This Wikipedia page gives a broad starting overview of speech-to-text concepts and is suitable for beginners who want orientation before deeper reading.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Speech-to-text (opens in a new tab)YouTube search
This YouTube search results page offers many video explanations of speech-to-text and suits readers who learn better by comparing multiple demonstrations, noting that it is a search page rather than one vetted lesson.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Speech-to-text (opens in a new tab)Reddit search
This Reddit search results page shows community questions and experiences about speech-to-text and suits readers who want informal discussion, noting that it is a search page and posts vary in reliability.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
Speech-to-text gives you a fast text draft from audio, but a transcript is only trustworthy after checking the parts that matter.