What is speech-to-text? Explained with a courtroom stenographer
Speech-to-text is AI that listens to spoken audio and writes down what was said as searchable text. Think of a courtroom stenographer typing a record of everything said.
- Author
Umair Mateen- Published
- Reading time
- 5 min
- Category
- AI
In one sentence: Speech-to-text is AI that listens to spoken audio, from a live call or a recording, and writes down what was said as text that people and software can search, read and use.
The everyday picture
Picture a stenographer in a courtroom. People speak, sometimes fast, sometimes over each other. The stenographer listens and types every word, so there is a written record of the whole hearing.
A good stenographer also notes who said each line, adds the full stops and commas, and learns the special words of the case in advance, like the names of the people involved. Later, anyone can read the record or search it, instead of listening to hours of talk.
Speech-to-text (also called speech recognition, or ASR for "automatic speech recognition") is that stenographer, done by software. It takes in sound and hands back a typed record.
The speech-to-text model listens like a stenographer: it turns the sounds into the most likely words, adds speaker labels and punctuation, and hands a typed record to your software.
How it works, step by step
- The audio comes in (people speaking in the room). It can be a live stream, such as a phone call, or a file you already recorded.
- The sound is cleaned and broken down (the stenographer tuning out the noise in the room). AWS describes the system cleaning up volume and background noise, then splitting the sound into tiny speech sounds.
- A trained model turns sounds into words (the stenographer's trained ear). The model learned from huge amounts of recorded speech. It picks the most likely string of words, using the words around each one to tell apart sound-alikes such as "there" and "their".
- The record gets tidied (the stenographer marking who spoke and adding punctuation). Services can add punctuation, write numbers as digits, give each word a time stamp, and label speakers as "Speaker 1" or "Guest 2". Microsoft calls this last part speaker diarization.
- Your software uses the text (the clerk filing the record). It can become meeting notes, subtitles, a searchable call log or the input to another AI tool.
Try it · Myth or fact
Speech-to-text: myth or fact?
Decide for yourself first, then tap each statement to check it.
0 of 4 checked
Where the picture stops
- It does not understand the case. A stenographer knows what is being discussed. The model only picks the words that most likely match the sounds, so it can write a confident but wrong word.
- It cannot ask anyone to repeat themselves. When people talk over each other or sit far from the microphone, it guesses. Microsoft lists overlapping speech, noise, distance from the microphone, accents and unusual words as common causes of errors.
- "Speaker 1" is not a name. Microsoft says its speaker labels show who spoke when, but are not designed to identify who a person is. A person still has to match labels to names.
Why it matters for your business
- Speed. Meetings and calls become text you can search in minutes, instead of being replayed or typed up by hand.
- Customer insight. AWS and Microsoft both list call centres as a main use: turning recorded calls into text so teams can spot common questions and problems.
- Access. Live captions and subtitles help people who are deaf or hard of hearing, and anyone watching with the sound off.
- Risk. Recording people has legal rules. Microsoft advises getting consent before you record and checking local laws on recording calls. Some services, such as Amazon Transcribe, can hide personal details like phone numbers in the transcript.
An example
Imagine a 15-person property management office that takes tenant calls all day. Staff write short notes after each call, and details get lost. With speech-to-text, callers hear that the call is recorded, and each call is transcribed with "Agent" and "Caller" labels. The office adds a list of its street and building names so the system spells them right. A manager reads the transcripts of any complaint calls before replying, rather than trusting a summary alone.
Try it · What would you do?
Your office wants to transcribe every customer call. What first?
A small office wants every customer call turned into text, so staff can search past conversations and stop typing notes by hand.
Pick what you would do. You can change your mind.
Frequently asked questions
How accurate is speech-to-text?
It depends on the audio. Clear speech from a good microphone in a quiet room does best; noise, crosstalk, strong accents and specialist words do worse. Accuracy is usually measured as word error rate: the share of words the system got wrong, missed or added.
What is the difference between real-time and batch transcription?
Real-time (streaming) transcription writes the text while people are still talking, for live captions or calls. Batch transcription works on files you already have and returns the full transcript afterwards.
Can it learn my industry's words?
Often, yes. Microsoft offers phrase lists and custom models, and AWS offers custom vocabularies, so product names and jargon are written correctly.
Is speech-to-text the same as a voice assistant?
No. Speech-to-text is the listening and writing part. A voice assistant also has to work out what you want and reply, often using a large language model and text-to-speech.
Related explainers
The DevDrift builds AI automation that turns calls and documents into useful records, with a person approving what matters: see how.
Sources
Could you explain this to a colleague now?
Mark it and posts that mention it will show a tick. Saved in this browser only.