ASR (Automatic Speech Recognition)
Technology that converts spoken language into written text automatically. Modern ASR systems such as Whisper use deep neural networks trained on large amounts of audio.
Whisper
An open-source speech recognition model released by OpenAI in 2022, trained on about 680,000 hours of multilingual audio. It supports 99 languages and can translate speech into English. More about this.
WER (Word Error Rate)
The standard accuracy metric for transcription: substitutions, deletions and insertions divided by the number of words in a reference transcript. A WER of 5% means about 95% of words are correct.
Speaker diarization
Working out who spoke when in a recording and labelling each segment by speaker. Our editor lets you assign speaker labels with one click per line.
Timestamp
A time code (for example 00:01:23) that marks where in the recording a piece of text was spoken. Timestamps let you jump straight to a quote.
Verbatim transcription
A transcript that records every utterance exactly, including filler words ("um", "uh"), false starts and repetitions. Used in legal and discourse-analysis work.
Clean verbatim
A transcript that keeps the speaker’s meaning and wording but removes fillers and stutters for readability. AI models like Whisper naturally produce clean verbatim text.
Intelligent verbatim
A lightly edited transcript that also fixes grammar and removes off-topic asides while preserving meaning. Common for business and content repurposing.
SRT (SubRip)
The most widely supported subtitle file format: numbered cues, start and end times with a comma before milliseconds, and the caption text. More about this.
VTT (WebVTT)
The web’s native caption format, starting with a WEBVTT header. It supports styling, positioning and speaker voice tags and works with HTML5 video. More about this.
Closed captions
Text versions of the dialogue and important sounds in a video that viewers can switch on or off. Essential for accessibility. More about this.
Open captions
Captions burned permanently into the video image, so they cannot be turned off. Common on TikTok and Reels.
Subtitles
On-screen text of the dialogue, often translated into the viewer’s language. Assumes the viewer can hear the audio. More about this.
Speech translation
Turning speech in one language directly into text in another language. Whisper translates speech from many languages into English in a single step. More about this.
Sample rate
How many times per second audio is measured, for example 44,100 Hz for CD quality. Speech recognition models like Whisper work at 16,000 Hz and resample automatically.
Bitrate
The amount of data per second of audio, such as 128 kbps. Higher bitrates keep more detail; speech stays very intelligible even at 64 kbps.
Codec
The method used to compress audio or video, such as MP3, AAC or Opus. The codec, not the file extension, decides whether a browser can play a file. More about this.
Container format
A file format that wraps audio, video and metadata together, such as MP4, MKV or WebM. A container can hold different codecs.
Lossless audio
Compression that keeps every detail of the original recording, as in FLAC or WAV. Ideal for archiving interviews. More about this.
Hallucination (ASR)
When a speech model outputs text that was never spoken, often during silence or music. Silence detection and filtering reduce it.
Punctuation restoration
Automatically adding commas, periods and capitalization to recognized speech. Whisper does this as part of transcription.
Code-switching
Changing between languages mid-conversation, such as Hindi and English (Hinglish). Multilingual models handle it better than single-language systems.
WebGPU
A modern browser technology that lets websites use the computer’s graphics chip for fast computation. It makes in-browser AI transcription much faster.
WebAssembly (WASM)
A browser technology that runs compiled code at near-native speed. It is the fallback engine for running AI models where WebGPU is unavailable.
On-device AI
AI that runs on your own computer or phone instead of a remote server. Your data stays private and there are no per-minute processing costs. More about this.
Real-time factor
How fast transcription runs compared with the audio length. "10× real time" means one hour of audio is processed in about six minutes.
Turnaround time
How long it takes to receive a finished transcript. Human services take hours or days; AI takes minutes.
Transcript
A written record of spoken words from an audio or video recording. More about this.
Dictation
Speaking text out loud so that it is written down, either by a person or by speech-to-text software. More about this.
Qualitative coding
Tagging passages of interview transcripts with themes during research analysis, often in tools like NVivo or ATLAS.ti. CSV exports make this easier. More about this.