Transcription

Transcription means converting spoken language into written text. In the past, people with headphones sat at desks typing along, for instance during court proceedings or interviews. Today, programs take over the task, receiving an audio recording and turning it into finished text. One hour of recording can thus often be put into writing in under a minute. An important distinction: transcription only writes down what was said. It does not translate into another language and does not summarize anything — those are separate steps that can be added afterward.

Why audio suddenly becomes usable as text

For a long time, an audio recording was a closed box for a computer. You could play it, but not search it. Anyone looking for a specific statement in a three-hour podcast had to listen to it. As text, the same search takes a moment. Transcription thus turns dead material into something you can actually work with.

This has given rise to an entire market. Journalists transcribe interviews, doctors dictate findings, companies automatically log video conferences. The time savings are considerable: a human typically needs three to five hours to cleanly transcribe one hour of conversation. This exact calculation explains why transcription features are now built into almost every conferencing program.

There is also an accessibility aspect. Subtitles in videos almost always originate from a transcription. For deaf and hard-of-hearing people, this determines whether a video is usable at all. In many countries, laws even mandate such subtitles for public providers.

From sound signal to sentence

A microphone initially only delivers a curve: air pressure fluctuations, thousands of measurements per second. The program breaks this curve into very short segments, usually a few hundredths of a second long. For each segment, it calculates which pitches occur and how strongly. Language must then be derived from this sequence of numbers.

This step is handled by a neural network, a program that has learned from a very large number of examples. It is trained with thousands of hours of audio for which the correct text is already available. In the process, it learns which sound patterns correspond to which sounds and words. Modern systems like OpenAI’s Whisper can thus handle several dozen languages with a single model.

A crucial trick is looking at context. The words “Lehre” (teaching) and “Leere” (emptiness) sound identical in German — only the sentence reveals the correct spelling. That’s why good systems always evaluate entire word sequences rather than individual sounds. A common misconception, by the way, is equating transcription with understanding content. The model recognizes sound patterns very reliably without grasping meaning — which is why errors continue to occur with technical terms, proper names, and strong dialects.

Subtitles, meeting minutes, and voice assistants

You most often encounter transcription on video platforms. YouTube’s automatic subtitles are created exactly this way, as are live subtitles in video conferences. Voice messages in messaging apps can now also be displayed as text. Anyone sitting on a train without headphones can simply read the message.

Voice assistants like Siri or Alexa begin every request with a transcription. Only afterward does a second system decide what to do with the text. Meeting tools work similarly: they transcribe the conversation and then generate a summary with a task list from the text.

In the news, the term usually comes up in the context of data protection. A transcription makes a conversation permanently searchable, and many providers process the recording on their own servers. In Germany, all participants in a video conference must therefore be informed before a transcript is being made. A second point of contention is errors in sensitive areas: a mistranscribed medication name in a medical report can cause real harm.