A plain-language glossary of transcription and subtitle terms

Which terms describe how speech becomes text?

Automatic speech recognition (ASR) is the technology that turns spoken audio into written text. In everyday conversation, people usually call it speech-to-text. Transcription is the wider job. A transcript is the finished written record, which often takes the recognized words and adds speaker labels, timing, and punctuation.

A model is the trained system doing the recognition, and training data is the collection of audio and text it learned from. Within the system, an acoustic model matches sounds to units of speech, while a language model scores how likely certain word sequences are so the recognizer picks a sensible phrase over an odd one. Newer systems often merge these two jobs into one model.

Decoding is the step where the system searches for the most likely sequence of words for an audio clip. The system also gives its own guess a number called a confidence score, which is a rough hint rather than a guarantee. Sometimes during silence or bad audio, a system produces a hallucination, meaning fluent text that nobody actually said. That is one reason a person still needs to review automatic transcripts.

Which terms describe accuracy?

Word error rate (WER) checks a transcript against a reference transcript. You count the words that were substituted (a wrong word in place of a right one), deleted (a missing word), or inserted (an added word), and divide that total by the word count of the reference. Character error rate (CER) works the same way by counting characters instead of words. It is used for languages and scripts where word boundaries are unclear.

A reference transcript is the verified, accurate text you measure a result against. Before counting errors, normalization cleans up both texts for a fair test, such as stripping out punctuation, matching up number formats, and standardizing spelling. Because of this, a WER score only means something if you also know the reference transcript, the normalization rules, and the test audio used.

WER counts every word the same, so missing the word 'the' hurts the score just as much as getting a surname wrong. Many people therefore check names, numbers, and specific terms separately. We do not publish a percentage here because scores change with the audio quality and the language.

Which terms describe speakers?

Diarization figures out who spoke when and breaks a recording into turns by speaker, without actually recognizing the words. A turn is one uninterrupted stretch of speech from one person. A speaker label is the tag put on that turn, like 'Speaker 1' or a person's name.

Overlapping speech is when two or more people talk at once, which makes both word recognition and diarization much harder. Voice activity detection (VAD) finds the parts of a recording where someone is talking and skips the silence and noise. To score speaker separation, diarization error rate (DER) measures how much time was given to the wrong speaker, missed altogether, or falsely flagged as speech.

Which terms describe languages and translation?

Language identification figures out which language someone is speaking. Code-switching is when a speaker switches between languages inside a conversation or even a single sentence, like mixing Hebrew and English. A dialect is a regional or social variety of a language. In Arabic, for example, Modern Standard Arabic is used for formal writing and broadcasts, while everyday spoken Arabic is mostly dialect.

Translation carries the meaning of words from one language into another. Transliteration swaps the script to match the sound, such as spelling an Arabic name with Hebrew letters. A right-to-left (RTL) language like Hebrew or Arabic is written from the right side of the page to the left, which changes how subtitles and documents lay out on screen when English words or numbers are mixed in.

KolWrite transcribes in over 100 languages, handles mixed speech like Hebrew and English, and translates between supported languages.

Which terms describe timing and writing style?

A timestamp marks when something was said, counting up from the start of the file. In video, a timecode labels a specific frame, usually written out as hours, minutes, seconds, and frames. Word-level timestamps give a start and end time for every single word. If you already have the text, forced alignment matches it against the audio to find those exact word times.

Verbatim transcription keeps every utterance on the recording, including filler words, false starts, and repetitions. Clean verbatim (sometimes called intelligent verbatim) cuts out the fillers and false starts while keeping the speaker's exact words and meaning. Punctuation restoration puts capital letters, commas, and periods onto the recognized words, and number formatting turns spoken numbers into digits.

Which terms describe files and captions?

SRT and VTT are plain-text subtitle files that pair lines of text with start and end times. VTT (or WebVTT) is the standard for web video and lets you control where the text sits and how it looks. For transcripts, TXT is plain text, DOCX is a Word document, and PDF is a fixed-layout file. JSON is a structured data format built for software, and it often holds word timings and speaker labels alongside the text.

Captions show speech and meaningful sounds for viewers who cannot hear. Subtitles transcribe or translate speech for viewers who can hear the audio but still need the text. Subtitles for the deaf and hard of hearing (SDH) is a subtitle style that includes sound descriptions and speaker identification. Closed captions sit on a separate track that can be turned on or off, while open (or burned-in) captions are drawn permanently into the video frame.

Which terms describe audio and workflow?

A codec is the method used to encode and compress audio. Lossless formats like WAV keep all the original audio detail, whereas lossy formats like MP3 throw away some data to make the file smaller. Bitrate is how much data gets processed for each second of audio; higher bitrates generally keep more detail. Sample rate is how many times a second the sound is measured. Mono audio uses a single channel, and stereo uses two.

Signal-to-noise ratio (SNR) compares how loud the voice is against the background noise. A low SNR is one of the main reasons speech recognizers make mistakes. On the workflow side, batch processing runs recorded files as a group, while streaming (or real-time) recognition processes audio live as it comes in. KolWrite works from uploaded recordings and does not do live captions during a call.

A custom vocabulary, or glossary, is a list of names and specific terms you give the system so it recognizes them more reliably. An API is an interface that lets other software send audio in and get text back. Review, sometimes called human-in-the-loop, is when a person checks and fixes the automatic transcript.