What Is Audio Transcription? Definition, Types, and How It Works

Audio transcription is the process of converting recorded spoken words into written text. Someone (or a machine) listens to an audio file — an interview, meeting, podcast, lecture, voicemail, phone call — and writes down what was said. The output is a transcript: a text document that reproduces the spoken content, usually with timestamps and sometimes with speaker labels. Modern audio transcription is typically done by AI models (OpenAI Whisper, Deepgram, AssemblyAI) that convert audio to text in minutes, with roughly 90-95% accuracy on clean audio.

A transcript is the finished document — the written text of what was said. Transcription is the process of creating that document. "I need a transcription" and "I need a transcript" are used interchangeably in most contexts, but strictly: transcription is the action, transcript is the artifact. The transcription industry uses "transcription service" for the process and "transcript" for the file the customer receives.

A transcription service is a company or software that converts audio recordings into written text. Two main types: (1) human transcription services (Rev, Scribie, GoTranscript, TranscribeMe) — human transcriptionists listen to your audio and type the transcript, typically 12-48 hour turnaround at $1-2.50 per audio minute. (2) AI transcription services (VexaScribe, Otter, Descript, Trint, Sonix) — machine learning models transcribe your audio in minutes at a fraction of the cost, typically $0.10-0.60 per hour. Hybrid services combine AI first-pass with human review for accuracy plus speed.

Four common types. (1) Verbatim — every word, filler ("um", "uh"), false start, and non-speech sound preserved. Required for legal depositions, conversation analysis, linguistic research. (2) Clean verbatim (or "intelligent verbatim") — filler words and false starts removed, but every meaningful word preserved. The default for most business, journalism, and podcasting use cases. (3) Edited transcription — the transcript is lightly restructured for readability while preserving meaning. Common for publication-ready interview quotes. (4) Phonetic transcription — uses IPA symbols to represent the sounds of speech rather than standard spelling. Used in linguistics and language learning.

Modern AI transcription follows five steps. (1) Audio pre-processing — the audio is normalized, resampled (usually to 16 kHz mono for Whisper), and voice activity detection removes long silences. (2) Automatic speech recognition (ASR) — a machine learning model (Whisper Large-v3, Deepgram Nova-3, AssemblyAI Universal-2) predicts the sequence of words that produced the audio. (3) Punctuation and casing — a separate model adds sentence boundaries, capitalization, and punctuation. (4) Diarization (optional) — a speaker-labeling model (pyannote.audio) identifies who spoke each segment. (5) Post-processing — timestamps are attached to each word or segment; the transcript is exported as TXT, DOCX, SRT, VTT, or JSON. Human transcription skips ASR and does the whole process manually.

Depends on the method and audio quality. Human transcription — 95-99% on clean audio; industry services like Rev advertise 99%+. AI transcription — 90-95% on clean audio (Whisper Large-v3 hits around 92% average across languages on the Open ASR Leaderboard), 75-90% on noisy or multi-speaker audio. Proper nouns, technical terms, and accents drop accuracy 10-20 percentage points regardless of method. Auto-captions on YouTube run 75-85% English, lower on other languages. The FCC's broadcast accuracy standard (47 CFR §79.1) does not specify a numeric threshold but requires captions to "match the spoken words to the fullest extent possible."

Common cases. (1) Research — qualitative researchers transcribe interviews for coding in NVivo/ATLAS.ti/Dovetail. (2) Journalism — reporters transcribe source interviews to find quotable material and verify accuracy. (3) Content repurposing — podcasters and video creators turn episodes into blog posts, show notes, and social clips. (4) Meetings — participants get searchable minutes without a dedicated note-taker. (5) Accessibility — deaf and hard-of-hearing users need text alternatives to audio content (ADA, WCAG 2.1). (6) Legal — depositions, court hearings, and evidence require verbatim transcripts admissible in court. (7) Medical — clinical dictation, patient encounters, HIPAA-governed workflows. (8) Language learning — transcripts help learners read along with native audio.

Wide range. AI transcription: $0.10-0.60 per audio hour depending on the service. VexaScribe starts at 30 minutes free (no card) then $2/month for 200 minutes. Human transcription: $60-150 per audio hour typically ($1-2.50 per minute at standard rate). Rush human transcription at 6-24 hour turnaround: $150-300 per audio hour. Legal verbatim (court-admissible): $200-400 per audio hour. See how much does transcription cost for the full breakdown by service type, turnaround, and audio quality.

Very close, with subtle industry distinction. Speech-to-text (STT) typically refers to the underlying technology and API — the raw capability of converting audio waveforms to text tokens, exposed as a developer service (OpenAI Whisper API, Google Cloud Speech-to-Text, AWS Transcribe, Azure AI Speech). Audio transcription refers to the end-to-end workflow of producing a usable transcript document — often adding punctuation, diarization, formatting, timestamps, and export to TXT/DOCX/SRT on top of the raw STT output. Transcription services use STT under the hood but package it with the surrounding UX. Automatic Speech Recognition (ASR) is another synonym for STT, used more in academic and technical contexts.