What is transcription?

Human vs automatic transcription

Human transcription means a person listens and types. It handles crosstalk, accents and context well, but it is slow and priced per minute of audio. Automatic transcription, also called speech-to-text or ASR (automatic speech recognition), uses a trained model to do the same job in a fraction of the time.

Automatic transcripts come in two broad styles. A verbatim transcript keeps every “um,” false start and repetition. A clean transcript smooths those out for reading. Most modern models land somewhere between the two by default.

Whichever method you use, the result should be checked where accuracy really matters. A legal record or a published quote deserves a human pass, while a transcript used only to find a moment in a recording can tolerate a few mistakes, because you will be listening to the audio anyway once you land on it.

How automatic transcription works

The audio is resampled and split into short chunks. A speech model converts each chunk into a sequence of likely words, using both the sound and what words tend to follow each other. The output is stitched back together, punctuated, and usually given timestamps, either per sentence or per word. Extra passes can add speaker labels or detect the language.

Some models run in the cloud and some are small enough to run on a laptop. Running locally means the recording never leaves your machine, at the cost of using your own CPU or GPU.

Common errors and how to reduce them

Even good transcripts have predictable weak spots:

• Names and jargon. Unusual names, product names and acronyms get replaced with common words that sound similar. A custom vocabulary or names list helps.
• Poor audio. Room echo, music beds and distant microphones hurt accuracy more than accents do.
• Crosstalk. When people talk over each other, one voice usually wins and the other is lost.
• Invented text in silence. Some models occasionally produce words during long silences or music. Skimming those sections is worthwhile.

Why timestamps matter

A transcript without timestamps is a document. A transcript with timestamps is an index: every word points back to a moment in the recording, so you can search the text and jump straight to the audio. That is what makes transcription the foundation for searching video and audio. See word-level timestamps for more.

In MediaFind

MediaFind transcribes on your computer using Whisper models with word-level timestamps, and a names and terms list keeps people and products spelled right. On Apple Silicon you can switch to the whisper.cpp engine, which runs on the GPU. Transcripts become searchable and can be saved as Word, PDF, SRT, VTT or plain text. For a step-by-step walkthrough, read how to transcribe video offline on a Mac.

Frequently asked questions

What is the difference between transcription and captioning?

Transcription produces text of what was said. Captions are that text split into short, timed chunks that display on screen in sync with the video, usually in a format like SRT or WebVTT.

Can I transcribe audio without uploading it?

Yes. Several speech recognition models run on a normal laptop, so the audio can be transcribed offline without sending it to a service.

Is automatic transcription accurate enough for quotes?

It is a good starting point, but check any quote against the audio before publishing. Timestamps make that quick, because you can jump to the exact moment.