Transcription glossary
Short definitions of the terms you meet while choosing a transcription tool or reading its documentation.
Transcription is the written record of speech from a recording. A handful of terms circle around it that tool descriptions tend to use without explaining: ASR is automatic speech recognition, diarisation is speaker recognition, WER is a measure of error, and verbatim is a word-for-word record keeping fillers and repetitions.
Transcription
The written record of speech from a recording. In the context of AI tools it means automatically converting an audio or video file into text, usually with timestamps. Not to be confused with phonetic transcription or with transcription in biology, which are unrelated senses of the same word.
ASR (automatic speech recognition)
The technology that turns an audio signal into text. Modern ASR systems are models trained on large collections of recordings; they do not understand content but predict the most probable sequence of words. Hence the typical errors: names, domain terms and acronyms that were rare in the training data.
Diarisation (speaker segmentation)
Splitting a recording into segments by who is speaking and tagging them with labels. It is sometimes called speaker recognition, which misleads: that term normally means identifying or verifying a particular person, which is a separate technology. Diarisation answers “who said what”, not “who is this”. It fails most where people talk over each other.
WER (word error rate)
A measure of transcription quality: the share of wrong words against a reference transcript, counting insertions, deletions and substitutions. A WER of 5% means one error in twenty words. The figure depends on the recording, so quoted without conditions — language, speaker count, audio quality — it says little.
Timestamps
Information about which second of the recording a given fragment of text came from. They let you jump back to a specific point instead of scrubbing blindly, and they are what makes subtitles possible. Without them a transcript is just text rather than a tool for working with audio.
SRT
A simple subtitle file format: successive blocks hold a number, a time range and the text. It is understood by media players, editing software and video platforms, which is why SRT export is the standard route from a transcript to subtitles under a video.
Verbatim
A word-for-word record that keeps repetitions, filler words and unfinished sentences. Used where the exact course of speech matters — in qualitative research and legal material. Its opposite is a cleaned-up record, easier to read but altering the form of what was said.
Transcript redaction
Cleaning a transcript of filler words, repetitions and false starts so it reads like written text. The meaning stays and the form changes — which is why material that depends on literal wording is left unredacted.
Bitrate
The number of bits per second of recording, given in kbps. It decides file size: an hour at 128 kbps is about 58 MB, at 320 kbps about 144 MB. For transcription accuracy the clarity of the recording weighs more than the bitrate, though a very low one can still hurt.
Lossy and lossless formats
A lossy format (MP3, AAC, OGG) discards some audio information to shrink the file; a lossless one (WAV, FLAC) keeps everything and weighs many times more. At typical settings speech loses nothing an ASR model depends on, but heavy compression and very low bitrates can degrade the result.
Model hallucination
A passage the model produced although nothing in the recording corresponds to it. In transcription it happens most often in silence, in noise and at the ends of files. That is why any AI transcript needs checking wherever a decision rests on it.
Personal data in a recording
A recorded conversation almost always contains personal data — the voice itself is data, and the content adds names and details. That brings GDPR duties: a lawful basis, a processing agreement with the tool's provider, and a defined retention period.