Transcription and subtitle glossary

Browse by topic

• Transcription & ASR (6)
• Captions & subtitles (5)
• Formats (5)
• Speakers (4)
• Timing (6)
• Localization (2)
• Accessibility (3)
• Quality (7)
• Video & audio workflows (1)

A–Z index

Transcription & ASR

• automatic speech recognition — Automatic speech recognition (ASR) is software that maps an audio waveform to text without a human typing every word. Modern ASR systems are statistical or neural models trained on large collections of speech and transcripts.
• clean verbatim — Clean verbatim is a transcript style that preserves meaning and sentence structure but removes fillers, false starts, and obvious stutters. It is more readable than full verbatim and still closer to the speaker than a rewritten summary.
• full verbatim — Full verbatim (sometimes called true verbatim) records speech with fillers, stutters, repetitions, and often nonverbal sounds, following the client’s notation rules. It is used when how something was said is part of the evidence.
• speech recognition model — A speech recognition model is the trained statistical or neural system that maps audio features to tokens or words. Examples include the Whisper family of checkpoints. The model is not the same thing as the product UI around it.
• transcription — Transcription is the conversion of spoken words in audio or video into written text. A transcript may be plain text or include timestamps, speaker labels, and notes about inaudible or overlapping speech.
• verbatim transcription — Verbatim transcription records speech as spoken, including fillers, false starts, and repetitions, according to a stated style guide. “Verbatim” is not one universal standard — clients specify full verbatim or clean verbatim.

Captions & subtitles

• closed captions — Closed captions are timed text that can be turned on or off by the viewer. They encode the spoken words and, for accessibility captions, relevant non-speech sounds. They travel as a separate track, not painted into the picture.
• forced subtitles — Forced subtitles are cues that appear even when the viewer has subtitles turned off, typically for foreign-language lines, signs, or in-world text that the story requires everyone to read.
• open captions — Open captions are timed text burned into the video frames. Viewers cannot turn them off. They are also called hardcoded or burned-in captions.
• SDH — SDH means Subtitles for the Deaf and Hard of Hearing. An SDH track includes dialogue plus sound cues and speaker identification so viewers who cannot hear the audio still get that information.
• subtitles — Subtitles are timed text shown with a video, usually as a toggleable track. In many locales they translate dialogue. In accessibility contexts the same file family also carries captions for viewers who cannot hear the audio.

Formats

• caption file — A caption file is a timed-text document a player can load as a track — commonly SRT, WebVTT, TTML, or SCC. It stores cues, not the video pixels.
• SCC — SCC (Scenarist Closed Caption) is a file format that stores CEA-608 caption data as hexadecimal byte pairs with timecodes. It is common in North American broadcast and finishing workflows.
• SRT file — An SRT (SubRip) file is a plain-text subtitle format. Each cue has an index, a start and end timecode using comma milliseconds, one or two lines of text, and a blank line.
• TTML — TTML (Timed Text Markup Language) is a W3C XML format for timed text. Streaming and broadcast pipelines often use a constrained profile such as IMSC rather than unrestricted TTML.
• WebVTT — WebVTT (Web Video Text Tracks) is a W3C subtitle and caption format for HTML5 video. A file begins with WEBVTT and uses cues with start and end times, typically with dot milliseconds.

Speakers

• overlapping speech — Overlapping speech is two or more people talking at the same time. It is one of the hardest conditions for automatic speech recognition and for diarization.
• speaker diarization — Speaker diarization answers “who spoke when?” It segments audio into speaker turns and assigns labels such as Speaker 1 and Speaker 2. It does not, by itself, name those people.
• speaker identification — Speaker identification matches a voice to a known person or a previously enrolled profile. It is a biometric or lookup task, not just splitting a file into Speaker 1 and Speaker 2.
• speaker labels — Speaker labels are the names or codes written next to each turn in a transcript — “Alex:”, “Speaker 2:”, or “Interviewer:”. They come from diarization, a human, or both.

Timing

• drop-frame timecode — Drop-frame timecode is a numbering system for 29.97 fps (and 59.94) video. It skips certain frame numbers so the timecode display stays close to real elapsed time. It does not delete video frames.
• frame rate — Frame rate is how many video frames are displayed per second — 24, 23.976, 25, 29.97, 30, 50, 59.94, and others. Caption timing that assumes the wrong rate will drift.
• SMPTE timecode — SMPTE timecode is the broadcast family of address codes defined by SMPTE (notably ST 12). It counts hours, minutes, seconds, and frames, and exists in drop-frame and non-drop-frame varieties for NTSC-related rates.
• subtitle synchronization — Subtitle synchronization is the alignment of each cue’s start and end times with the spoken words (and any required shot changes). Out-of-sync cues appear too early, too late, or drift across the file.
• timecode — Timecode is a numeric label for a position in media, usually hours:minutes:seconds plus frames or milliseconds. Transcripts and subtitle files both use timecode, but they may use different precision.
• timestamp — A timestamp is a time marker placed in a transcript to show when a span of speech occurred. Intervals (every 30 seconds, every speaker change) are a client rule, not a property of ASR.

Localization

• localization — Localization adapts media for another language and market. For video that may mean translated subtitles, dubbed audio, or culturally adjusted on-screen text. It is broader than word-for-word translation.
• multilingual transcription — Multilingual transcription produces text for speech in more than one language — either one file per language, a language-id plus recognition pipeline, or a single transcript that follows code-switching.

Accessibility

• audio description — Audio description (AD) is spoken narration that explains important visual information during pauses in the dialogue, for viewers who are blind or have low vision. It is a separate access service from captions.
• caption accessibility — Caption accessibility means the timed text actually works for people who cannot hear the audio: accurate words, sync, completeness, useful sound cues, and a player that can display the track.
• WCAG — WCAG (Web Content Accessibility Guidelines) is the W3C standard for making web content more accessible. For video, Success Criterion 1.2.2 requires captions for prerecorded synchronized media, with related criteria for live captions and audio description.

Quality

• character error rate — Character error rate (CER) is the same edit-distance idea as WER, computed on characters instead of words. It is common for languages without clear word boundaries and as a complement to WER.
• confidence score — A confidence score is the recognizer’s own estimate that a word or segment is correct. It is not an independent accuracy measurement and is not comparable across vendors without a shared calibration study.
• hallucination — In speech recognition, a hallucination is invented text that does not correspond to the audio — for example, repeated phrases, leftover prompt language, or fluent sentences during silence or music.
• punctuation restoration — Punctuation restoration is the step that adds periods, commas, question marks, and often capitalization to raw ASR output that was predicted as an unpunctuated word stream.
• transcript QA — Transcript QA is the review pass after drafting — checking words, speakers, timestamps, tags, and style-guide rules before delivery. It is separate from the act of capturing speech.
• transcription accuracy — Transcription accuracy is how closely a transcript matches what was said, under a stated definition of “match.” Researchers usually report word error rate. Clients may mean names, numbers, and style-guide compliance instead.
• word error rate — Word error rate (WER) is the standard research metric for speech recognition. It counts substitutions, deletions, and insertions against a reference transcript, then divides by the number of words in that reference.

Video & audio workflows

• audio extraction — Audio extraction is pulling the soundtrack out of a video container so a speech recognizer can process it. The picture is not required for transcription, only the audio stream.

All published terms