Speech-to-text

What it means

Speech-to-text, also called automatic speech recognition, listens to an audio track and writes down what was said, usually with a timestamp for every word. Modern models are trained on huge amounts of speech in many languages and handle accents and casual talk far better than older systems.

In video work the transcript is a tool, not just a document. It powers auto captions, lets editors search hours of interviews for one phrase, and enables text-based editing where deleting a sentence in the transcript removes that part of the clip. It is also one input to multimodal video analysis, adding meaning that the picture alone cannot provide.

How it works in practice

Accuracy depends mostly on the recording. A talking-head clip with a lavalier mic transcribes almost perfectly; a helmet camera at speed may return fragments. Speak toward the camera when you want your words captured.

Editors such as Premiere Pro and DaVinci Resolve have built-in transcription, and open models like Whisper run locally. Set the right language before transcribing, since auto-detection can misfire on short clips.

Expect timestamps to be close but not frame-accurate. For captions that is fine; for tight cuts on dialogue, check the cut points against the waveform.

What to watch out for

Overlapping voices, shouting and wind noise are the classic failures. The model may invent plausible words to fill gaps, so a transcript can look confident and still be wrong.

Proper names, slang and sport jargon get mangled. A trick name or a local spot name often comes back as a similar common word. Keep a glossary of names and correct them in one pass.

Privacy matters too. Cloud transcription services receive your audio, which may include other people talking. For private recordings, prefer an offline model or check what the service does with uploaded files before you send anything.