Glossary
What is transcription?
By the KeyMo team · Updated
Transcription is the process of turning spoken audio into written text, done by a person, by software, or by software that a person then checks. The written result is a transcript. In biology the word means something different: copying a gene's DNA into RNA. This page covers the audio sense.
Three ways to transcribe
| Method | How it works | Strength | Weak spot |
|---|---|---|---|
| By hand | A person listens and types | Handles names, accents and context | Slow and costly for a long episode |
| Automatic (speech to text) | Software turns audio into words | Fast and cheap, good for a first pass | Misses names, jargon and overlapping speech |
| Automatic plus review | Software drafts, a person corrects | Fast and checked | Still needs someone to read it |
The speech to text entry covers the automatic side.
What affects accuracy
- Audio quality. Room echo, a distant microphone and music under speech all add errors.
- Crosstalk. When two people talk at once, one of them usually gets lost. See diarization for telling speakers apart.
- Accents and jargon. Product names, places and technical terms are the likeliest misses.
- Length and noise. YouTube's help page says its auto-sync option is not recommended for videos over an hour or with poor audio.
The human-check step
YouTube says automatic captions come from machine learning, that quality varies, and that they should always be reviewed and edited. The W3C accessibility guidance goes further: automatically generated text does not meet accessibility needs unless it is confirmed fully accurate, and it usually needs significant editing. Its example is that dropping one word, "not", can make the text say the opposite of the audio.
Style is a choice too. W3C guidance says legal depositions need verbatim text including "um" and "ah", while for most web content non-substantive filler may be left out. Unclear speech can be marked [unintelligible] instead of guessed.
An example from a two-host episode
Two hosts laugh over each other while a guest says a company name for the first time. The draft reads "we signed the deal with tree fork". A person who knows the show spots it and fixes the name, adds the speaker labels and removes a repeated word. That ten-minute pass is the difference between a draft and a document you can quote.
Common mistakes
- Assuming a clean-sounding result is a correct one. Wrong words can still read smoothly.
- Posting machine text without reading the numbers and names.
- Recording with a laptop microphone across the room, then blaming the software.
Where KeyMo touches it
KeyMo shows speech in its clips as word-by-word captions, in one of 12 styles or none. Treat them like any automatic transcription: read the captions before you post, especially names and numbers. The result of transcription on its own is a transcript.
Sources: W3C WAI, Captions/Subtitles (checked 11 October 2026); W3C WAI, Transcribing audio and video for captions (checked 11 October 2026); YouTube Help, Use automatic captioning (checked 11 October 2026); YouTube Help, Add subtitles and captions (auto-sync) (checked 11 October 2026).