Transcription glossary

Transcription

Converting spoken audio into written text. Automatic (AI/ASR) transcription is done by software in minutes; manual transcription is done by a human typist at roughly 4–6 hours of work per hour of audio.

ASR (Automatic Speech Recognition)

The machine-learning technology that converts speech to text — the engine inside every AI transcription tool. Modern ASR models are trained on millions of hours of speech across many languages.

WER (Word Error Rate)

The standard accuracy metric for transcription: the percentage of words that are wrong, missing or wrongly inserted versus a perfect reference. 5% WER means 95% accuracy. Lower is better; clear audio in major languages typically achieves 1–5% WER with modern ASR.

Speaker diarization

Automatically detecting who spoke when, and labeling segments by voice (Speaker 1, Speaker 2…). Diarization distinguishes voices within a recording — it does not identify who a person actually is.

Voice identification

Matching a voice to a specific person's identity — a biometric technology distinct from diarization. Trustample deliberately does not do voice identification.

Timestamps (word-level)

The start/end time recorded for each word during transcription. Word-level timestamps power click-to-replay editing, accurate subtitle cue timing, and search-to-audio navigation.

SRT (SubRip Subtitle)

The most widely supported subtitle file format: numbered cues with start/end timestamps (comma before milliseconds) and text. Accepted by YouTube, Premiere, DaVinci Resolve, CapCut, VLC and virtually everything else.

VTT (WebVTT)

The web-native subtitle format used by HTML5 video players. Nearly identical to SRT, but with a WEBVTT header, dots before milliseconds, optional cue settings and styling.

Verbatim vs clean read

Verbatim transcription keeps every um, false start and repetition; clean read (or intelligent verbatim) removes fillers and fixes fragments for readability. AI transcripts are near-verbatim by default; Trustample's AI cleanup action produces a clean-read version.

Code-switching

Alternating between languages within one conversation (e.g. Spanish–English, French–English). Word-level borrowing transcribes well; sentence-level switching is among the hardest problems in ASR.

Custom vocabulary (keyword boosting)

Supplying names, jargon and product terms to the ASR engine before transcription so it recognizes them correctly — the cheapest accuracy upgrade available for specialized audio.

Language detection

The engine's automatic guess at the spoken language. Reliable on clear, single-language audio; explicit language selection beats it on casual, quiet or mixed-language recordings.

Crosstalk (overlapping speech)

Multiple people speaking simultaneously. The single hardest condition for ASR and diarization — overlapping words typically transcribe as fragments from the loudest voice.

Characters per second (CPS)

Subtitle reading-speed metric: how many characters a viewer must read per second of display time. Comfortable adult reading speed is roughly 15–17 CPS; faster requires shortening the caption text.

The 42-character rule

The industry guideline (popularized by Netflix's style guide) keeping each subtitle line at 42 characters or fewer for at-a-glance readability.

Paranoid mode

Trustample's strictest privacy setting: the uploaded audio file is deleted permanently the moment transcription succeeds, leaving only the text.

Prompt caching

An efficiency technique for AI features on long transcripts: the transcript is cached at the AI provider so follow-up questions don't re-process the whole text — faster answers, lower cost.

Merchant of record (MoR)

A payment company (like Paddle) that legally sells the product to you, handling sales tax/VAT collection and remittance worldwide. It's why Trustample prices can include taxes.

Put the vocabulary to work: transcribe a recording free or read how accurate AI transcription really is.