Skip to content

Glossary

The conversion step, and nothing the term promises beyond it

Speech-to-text is any system that turns spoken audio into written words. What the term covers, what it does not, and where on-device and cloud approaches differ

· 3 min read

On this page (4)
  1. What the term does and does not include
  2. Where the conversion actually happens
  3. Where dictation software fits into this
  4. What accuracy actually measures

Speech-to-text is any system that takes spoken audio and produces a written transcript of it — nothing more is implied by the term itself. It says nothing about where the audio is processed, what happens to the words afterward, or whether anything downstream understands what was said. A subtitle generator, a voicemail transcript, a courtroom stenography tool and a dictation app are all speech-to-text; what separates them is everything built around that one conversion.

What the term does and does not include

Speech-to-text is the conversion step alone. It is easy to conflate with things that usually sit next to it:

  • Not the same as a voice assistant. An assistant takes the transcribed text and does something with it — answers a question, runs a search, sets a timer. Speech-to-text on its own just produces the text; whether anything acts on it is a separate layer, present in some products and absent in others.
  • Not the same as voice commands. A tool that recognizes "delete that paragraph" as an instruction is doing command recognition, matching speech against a fixed set of actions rather than transcribing it as prose. Some products do both; the two are different problems.
  • Not the same as translation. Converting spoken French into written French is speech-to-text. Converting it into written English is translation layered on top, and requires a second step the plain conversion does not include.

Where the conversion actually happens

The audio can be processed two ways, and the difference matters more than it sounds:

On-device — the engine runs on the same machine that recorded the audio, and the words never leave it. This works with no network connection at all, and nothing about the audio or the transcript is uploaded anywhere.

Cloud-based — the audio is sent to a server, converted there, and the text is sent back. This usually draws on a larger model and can support more languages, but it needs a connection for every sentence, and the audio reaches a server it did not start on.

Neither approach is universally better; they trade capability for where the audio goes, and the right one depends on what a specific use actually needs. The honest comparison between Coii VoiceInput and Wispr Flow covers one on-device tool set against one that is not, including what each trade actually costs.

Where dictation software fits into this

A dictation app is a specific, narrower use of speech-to-text: converting speech into text and typing it directly into whatever application already has focus, rather than producing a standalone transcript file to be opened separately. What is dictation software covers that distinction in full.

What accuracy actually measures

Speech-to-text systems are commonly compared by how often the transcript matches what was actually said. A lower error rate means fewer words came out wrong, but the number alone says nothing about where the audio went to get that result — a private on-device engine and a cloud service can both publish a low error rate while differing completely on what happens to the audio itself. The broader case for speech-to-text on a Mac covers what that tradeoff looks like applied to one operating system specifically.

Questions

Is speech-to-text the same thing as a voice assistant?
No. Speech-to-text only converts speech into text. A voice assistant adds a layer on top that understands the text and acts on it — answering, searching, controlling an app. Plenty of speech-to-text tools do nothing beyond the conversion.
Does speech-to-text always need an internet connection?
No — it depends on where the conversion happens. Some tools run the engine on the device itself and need no connection; others send audio to a server and need one for every sentence. Both are speech-to-text; only one works on a plane.