Glossary: Meetings

What is Speech-to-text model?

The AI that turns spoken audio into written text. It can run on a company's servers or on your own computer.

Updated Nov 1, 2026By Bruce W.

A speech-to-text model (also called automatic speech recognition, or ASR) is the AI that listens to audio and writes down the words. It’s what powers meeting transcripts, voice typing and captions.

Local or cloud

  • Cloud models run on a provider’s servers. Your audio is sent to them over the internet.
  • Local models run on your own computer. The audio never leaves it, and they work offline.

Open models like OpenAI’s Whisper and NVIDIA’s Parakeet made good local transcription practical on ordinary laptops. More in Local vs cloud transcription.

How accuracy is measured

Usually by word error rate: wrong words per 100 on a public test. Real meetings, with accents, crosstalk and bad microphones, are harder than tests, so trying a model on your own recordings tells you more.

An example

Picture a 30-minute call recorded on a laptop. A small, fast model can write a live transcript as people talk, with a few slips like “cell” for “sell”. After the call, a larger, slower model can go over the same audio and fix most of those slips. That’s why some apps use two models: one for speed, one for accuracy.

Choosing a model

  • Size: bigger models are usually more accurate, but slower and hungrier for memory.
  • Languages: some models are English-only, others handle dozens of languages.
  • Your hardware: a graphics card lets you run bigger models comfortably.
  • Your audio: accents, jargon and noisy rooms affect models differently, so test on a real recording.
  • Where it runs: for private conversations, a model on your own computer keeps the audio there. See local transcription.

In Fetch

Fetch runs speech-to-text on your PC with two models: a fast live one during the call, and a careful clean-up pass after. You can pick from Parakeet, Whisper, IBM Granite Speech and others, and compare them on the same meeting. See Choose a speech model.

Questions people ask

What is word error rate?

The usual accuracy measure for speech-to-text. It counts wrong, missing and extra words per 100 words on a test recording. Lower is better.

Can a speech-to-text model handle more than one language?

Many can. Multilingual models like Whisper handle dozens of languages, though accuracy is usually best in English. Some models are English-only but faster, so check before you choose.

Still scratching your head?

Send us a note and a real person will get back to you.

Contact support