What is transcription? Definition, types and the best tools

How spoken words become text, which transcript style you need, what accuracy really means, and what to use.

Transcription is turning speech from an audio or video recording into written text. It's done by a person typing what they hear, by speech recognition software, or by software first and a person checking after.

The word has another meaning in biology (copying DNA into RNA, covered in the FAQ). This page is about the audio kind: how a transcript gets made, the styles you can order, how to judge accuracy, what it costs, and the tools we compared, from the free open-source model to the human service at $1.99 a minute.

Romain CochardCEO of Hack'celeration
Updated 20266tools compared4transcript styles explained8common questions answered

Some links to tools may be affiliate links. The definition and our take on each tool stay the same either way.

How it works · From sound to text

How transcription works: three ways to get a transcript

Every transcript starts with the same input, a recording (a meeting, an interview, a podcast, a lecture, a voicemail), and ends with the same output, a text file you can read, search, quote and edit. What changes is who does the listening.

1. Human transcription. A transcriptionist plays the recording, types what they hear, rewinds the hard parts and marks who is speaking. It's the slowest route: a trained typist spends several hours on one hour of audio, longer when the sound is poor or people talk over each other. It's also the one that handles accents, jargon, crosstalk and bad microphones best.

2. Automatic transcription (ASR). Automatic speech recognition software cuts the audio into tiny slices, turns each slice into likely sounds, then uses a language model to pick the most probable words and add punctuation. Modern engines also do speaker diarization ("Speaker 1, Speaker 2") and timestamps. An hour of audio comes back in minutes, for well under a dollar on a developer API and $10 on Sonix's pay-as-you-go rate.

3. Hybrid. Software produces the draft, a person corrects it. Many "human-verified" services work this way now, and it's what you end up doing yourself when you clean up an AI transcript before publishing it.

Two more distinctions matter when you pick a tool. Live versus file: live transcription writes as people speak (a meeting bot, live captions), file transcription works on a recording you upload, and the quotas are often counted separately. Transcription versus translation: a transcript stays in the language spoken; turning it into another language is a second step, sometimes billed again.

Before you choose, decide three things: who reads the result (you, a client, a court, a search engine), how many hours you have each month, and whether the audio can leave your company (health, legal and HR recordings often can't go to just any cloud service).

Then check five things on any tool before you compare prices: languages (Otter 17, some in beta; Descript 25, Sonix 54+), speaker labels and timestamps in the export, the export formats on the plan you'll buy, not only on the top one, custom vocabulary for names and jargon, and what happens to your audio: where it's stored, for how long, and whether it trains models.

Transcript styles · Verbatim, clean, edited

Types of transcription: which style to order

Ask for a transcript and the first question a service asks back is the style. Same recording, very different text. Here's one sentence in each style:

  • Full (strict) verbatim: "So, um, we, we were thinking [laughs] that the, uh, launch could move to May." Every filler, false start, stutter and audible sound is kept. It's what legal work and conversation research need, because a hesitation can carry meaning.
  • Clean (intelligent) verbatim: "We were thinking that the launch could move to May." Fillers and repetitions go, the words stay the speaker's. It's the default for interviews, meetings, podcasts and most business uses.
  • Edited: "We plan to move the launch to May." Grammar is fixed and wording tightened for reading. It suits published articles, reports and blog posts made from a talk, not quotes you'll attribute to someone.
  • Phonetic: how the words sound, written in a phonetic alphabet. It's a niche for linguists and speech therapists, written in the International Phonetic Alphabet.

AI tools give you something between full and clean verbatim by default: the words spoken, with punctuation and speaker labels, and some fillers dropped. Several let you choose (Rev and Sonix have a verbatim option, Descript can remove filler words in one click). If you order from a human service, the style changes the price, and full verbatim usually costs more because it takes longer.

Format is a separate choice from style. A plain text or Word file is for reading. An SRT or VTT file carries timestamps per line so a video player can show it as captions. A legal-format transcript has numbered pages and lines so a lawyer can cite "page 12, line 4", and it's often priced by the page instead of the minute (Rev charges $2.25 a page for a legal rough draft).

Our tip: always ask for timestamps, even in a text file. They cost nothing on AI tools, and when a quote looks wrong you'll find the spot in the audio in seconds instead of scrubbing for ten minutes.

Accuracy and cost · WER, quotas, hidden limits

Transcription accuracy and cost: how to read the numbers

Vendors love to print "up to 99% accuracy". The number behind that claim is usually the word error rate (WER): the substitutions, deletions and insertions in the transcript, added up and divided by the number of words actually spoken. A 5% WER means 5 errors per 100 words, about one mistake every two sentences. Accuracy is simply 100% minus the WER.

Two catches. First, the WER a vendor publishes was measured on its own clean test audio. Your call recorded on a laptop with two people talking over each other will score worse, sometimes much worse. Second, WER counts every word the same: a missed "the" and a wrong drug name each count as one error. That's why the accuracy figure to trust most is the one you measure yourself: transcribe 10 minutes of your typical audio on two tools, correct it, and count.

Cost depends on how the tool counts. Here are the units you'll see:

  • Per audio minute (human services, developer APIs): you pay for the length of the recording. One hour of human transcription at Rev's $1.99 a minute is $119.40; the same hour through OpenAI's gpt-transcribe API at $0.0045 a minute is $0.27.
  • Per hour of media (Sonix, Descript): your plan includes so many hours a month; Sonix bills extra hours at $10 each.
  • Minutes per month plus file caps (Otter, Notta): a monthly allowance of transcription minutes, and a separate cap on how many files you can upload or how long one conversation can run.

Three limits we found while comparing, that roundups skip:

  1. Otter's free plan allows 3 file imports per user, in total, not per month, and stops each conversation at 30 minutes; Pro allows 10 imports a month and 90 minutes per conversation. If you mostly upload recordings rather than record meetings live, you'll hit those walls first.
  2. Notta's free plan cuts each conversation at 3 minutes, even though it gives you 120 minutes a month, and exporting transcripts starts on Pro. It's enough to try the tool, not to transcribe an interview.
  3. On Sonix, an extra seat ($25 a month) doesn't add hours: the hours belong to the workspace, shared by everyone. A team of four on Core still shares 5 hours a month.

What a free plan really gives you. Otter offers 300 minutes a month, Notta 120, Descript 60 media minutes and Rev 45 AI minutes. Whisper has no monthly cap, because it runs on your own computer. Each free plan is a trial of the workflow, not a way to clear a backlog.

Privacy is part of the cost too. Ask where audio is stored, for how long, and whether it's used to train models. For health data in the US, you need a signed BAA (business associate agreement under HIPAA), and it's rarely on the cheap plans: Otter, for instance, lists HIPAA compliance as an add-on of its Enterprise plan, not of Pro or Business.

What it's for

Who needs transcription, and which kind

AI transcription is enough if…

  • You record meetings and want notes, action items and search, not a perfect record (see the best AI meeting minutes tools)
  • You run a podcast or YouTube channel and need captions you'll skim before publishing (pair it with the best podcast software)
  • You interview people for articles and check each quote against the audio anyway
  • You're a student or researcher with clear recordings and more time than budget
  • You build software and need speech to text inside your own app (an API)

Pay for a human pass if…

  • The transcript is evidence: depositions, hearings, police interviews
  • Medical dictation or patient conversations, where one wrong word is a safety issue
  • Research coding that depends on pauses, overlaps and hesitations (full verbatim)
  • Heavy accents, poor sound or many speakers talking over each other
  • Captions must meet an accessibility standard and you can't proofread them yourself
Related terms

Transcription vs captions, subtitles, dictation and translation: what changes

These words get used interchangeably. They're not the same deliverable, and they're not always priced the same.

TranscriptionNeighbor terms
What you getThe spoken words as a text documentCaptions: timed text shown on a video
Subtitles: timed text, usually in another language
TimingOptional timestampsCaptions and subtitles: every line is timed (SRT, VTT)
LanguageSame language as the audioTranslation: a second language, a separate step
Subtitles: often translated
Who speaksAny recording, any number of speakersDictation: one person speaking to be written down, often live
Includes soundsOnly in full verbatimClosed captions: also [music], [laughter] for deaf viewers
Typical fileTXT, DOCX, PDFCaptions: SRT, VTT
Dictation: text typed into your app
Typical toolsOtter, Rev, Sonix, Notta, WhisperCaptions: Descript, Rev
Dictation: your phone's voice typing

Open captions are burned into the video and always visible; closed captions can be switched on and off by the viewer.

The tools

The best tools for transcription, by type of need

We compared six tools, one per way of working. The quick route: meetings, Otter; files that may need a human, Rev; a steady flow of uploads, Sonix; podcasts and video, Descript; the smallest budget, Notta; private audio or developer use, Whisper. Fireflies.ai works much like Otter for meetings (Pro from $10 per seat a month billed annually), so we cover it with the best AI meeting note takers. For the full ranking of AI engines, see the best AI transcription software; for captions, the best subtitle generators.

01

Otter: live transcription for meetings

Meetings

Otter joins your Zoom, Google Meet or Teams calls, writes the conversation live with speaker names, then adds a summary and action items. You can also record in person from the mobile app.

Who it's for. People whose audio is mostly meetings and who want searchable notes more than a perfect record.

The limit. It's built for live meetings, not for uploads: the free plan allows 3 file imports per user in total and stops each conversation at 30 minutes; Pro allows 10 imports a month. Transcription covers 17 languages, some still in beta; only six (English, Spanish, French, German, Japanese and Chinese) are listed with the plans, against 54+ on Sonix.

The tip. Annual billing roughly halves the bill: Pro is $8.33 per user a month billed annually, against $16.99 month to month. Students and teachers with a .edu email get a further 20% off, which brings Pro to $6.67 a month billed $79.99 a year.

Moving on? See the best Otter alternatives.

Price
Free (300 minutes a month); Pro $8.33 per user/month billed annually ($16.99 monthly); Business $19.99 per user/month billed annually
Best for
Live meeting notes
Limit
Free: 3 file imports in total, 30 minutes per conversation; Pro: 10 imports a month
Try Otter →
02

Rev: AI transcription with a human option

Human-verified

Rev does both kinds: AI transcripts and captions on a subscription, and human transcription ordered per minute, guaranteed 99%+ accurate, when the stakes are high. The two live in the same account, so you can send only the hard files to a person.

Who it's for. Journalists, lawyers, researchers and anyone who sometimes needs a transcript they can stand behind.

The limit. Human work adds up fast: $1.99 a minute is $119.40 for a one-hour interview. Legal formatted transcripts are priced by the page ($2.25 for a rough draft, $2.50 for a premium transcript) and aren't available on a subscription.

The tip. A paid plan cuts the human rate: 3% on Essentials monthly, 10% on Essentials annual, up to 15% on Pro annual. If you order human transcripts every month, the subscription can pay for itself on the discount alone.

Price
Free (45 AI minutes a month); Essentials $25.49 per seat/month billed $305.90 yearly; human transcription from $1.99/minute
Best for
Files that need a human check
Limit
Human and legal transcripts get expensive
Try Rev →
03

Sonix: pay by the hour for uploaded files

Files

Sonix is built around uploaded recordings: you drop in interviews, lectures or podcasts, get a timestamped transcript in an in-browser editor, and export to Word, PDF, SRT or VTT.

Who it's for. Researchers, producers and small teams with a steady flow of files rather than live meetings.

The limit. Hours belong to the workspace, not to each person: an extra seat costs $25 a month and adds no hours. A team that grows has to move up a plan, not just add seats.

The tip. Pay as you go costs $10 an hour with no subscription. Below 2.5 hours a month that costs less than Core ($25 a month for 5 hours), and extra hours on any plan are billed at the same $10.

Price
Pay as you go $10/hour; Core $25/month (5 hours) or $275/year; Advanced $50/month (20 hours); Pro $80/month (40 hours)
Best for
Steady volumes of uploaded files
Limit
Extra seats don't add hours
Try Sonix →
04

Descript: edit audio and video by editing the transcript

Editing

Descript transcribes your recording, then lets you cut the audio or video by deleting words in the text. Remove a sentence from the transcript and it's gone from the recording. Filler words can be removed in one pass.

Who it's for. Podcasters, video creators and marketers who need a transcript as a means to edit and publish, not as the final deliverable.

The limit. Plans count media hours, not transcription alone: the free plan gives 60 minutes a month. If all you need is text, you pay for an editor you won't use.

The tip. Annual billing saves up to 35%: Hobbyist is $16 a month billed annually versus $24 month to month, for 10 media hours a month.

Full review: Descript, and the plan-by-plan breakdown in Descript pricing.

Price
Free (60 minutes a month); Hobbyist $16/month billed annually ($24 monthly); Creator $24 ($35); Business $50 ($65)
Best for
Podcast and video editing
Limit
Media hours shared with editing
Our rating
3.4/5
Try Descript →
05

Notta: a cheap all-rounder for meetings and files

Budget

Notta records meetings, transcribes uploaded files and adds AI summaries, with an add-on for translation. It covers the same ground as Otter at a lower annual price.

Who it's for. Freelancers and small teams who want meetings and files in one tool for under $10 a month on annual billing.

The limit. The free plan cuts each conversation at 3 minutes, even with 120 minutes a month to spend, and doesn't export transcripts: export starts on Pro. You can try it, not use it for real work.

The tip. Pro is $8.17 a month billed $97.99 a year, for 1,800 minutes a month and recordings up to 5 hours each. Pick the plan and the billing cycle carefully: monthly to annual is a switch in billing, but annual back to monthly, or Pro to Business, means cancelling and subscribing again.

Price
Free (120 minutes a month, 3 minutes per conversation); Pro $8.17/month billed $97.99 yearly; Business $16.67/month billed $199.99 yearly
Best for
Low-cost meetings and files
Limit
Free: 3 minutes per conversation, no transcript export
Try Notta →
06

Whisper: the free open-source model

Open source

Whisper is OpenAI's speech recognition model. Its code and model weights are released under the MIT License, so you can run it on your own computer or server for free, or call it through OpenAI's API.

Who it's for. Developers, and anyone with sensitive audio who wants it never to leave their own machine. Many desktop apps are built on it.

The limit. It's a model, not an app: no editor, no meeting bot, and no speaker labels out of the box. The large model needs about 10 GB of graphics memory; the turbo model about 6 GB.

The tip. Through the API, Whisper costs $0.006 a minute and gpt-4o-mini-transcribe $0.003, so an hour of audio is $0.36 or $0.18. For a backlog of hundreds of hours, that's the cheapest paid route on this page, but OpenAI has deprecated both and shuts them down in 2027; their replacement, gpt-transcribe, costs $0.0045 a minute.

Other options: Whisper alternatives.

Price
Free to self-host (MIT License); API $0.006/minute until its 2027 shutdown
Best for
Developers, private audio, big backlogs
Limit
No interface or speaker labels by default
See Whisper →

Is your tool missing from this page? Email us: hello@hackceleration.com

FAQ

Transcription FAQ

  • What is transcription in biology?

    In biology, transcription is the process of making an RNA copy of a gene's DNA sequence. That copy, messenger RNA (mRNA), carries the gene's protein information; in humans it moves from the cell nucleus to the cytoplasm, where it's used to build the protein. It shares only the idea of copying with audio transcription.

  • How long does it take to transcribe one hour of audio?

    By hand, several hours of typing for one hour of clear audio, and more with several speakers or poor sound. AI tools usually need a few minutes; Word's Transcribe warns an upload can take up to about the length of the file, depending on your connection. Human services take longest: Rev guarantees its human transcripts in 12 hours or less.

  • Can Microsoft Word transcribe audio?

    Yes, with a Microsoft 365 subscription: Word's Transcribe turns an uploaded .wav, .mp4, .m4a or .mp3 file into a transcript with speakers labeled and timestamps, up to 300 minutes of uploaded audio a month. On the web it works only in the new Microsoft Edge and in Chrome, so it's a handy free extra for occasional files if you already pay for Office.

  • Can ChatGPT transcribe audio?

    Yes: OpenAI sells its transcription models through its API, which is the route for long recordings. gpt-transcribe costs $0.0045 a minute, and gpt-4o-transcribe-diarize, which also labels who is speaking, $0.006 until OpenAI retires it in 2027. A chat window gives you no editor, no timestamps to click and no SRT export, so for a one-hour file a dedicated tool is easier.

  • Is transcription a good side job?

    It can be, if you type fast, because pay is per audio minute, not per hour worked. Rev, for example, lists $0.40 to $1.10 per audio or video minute for transcriptionists and $0.55 to $2.00+ for legal work. At those rates one hour of audio pays $24 to $66, however long it takes you: your typing speed decides your real hourly rate.

  • What file formats do transcription tools accept and export?

    Most accept MP3, WAV, M4A and MP4; Sonix takes 44+ audio and video formats. Exports are where plans differ: Otter's free plan exports only mp3 and txt, with pdf, docx and srt from Pro, while Sonix exports to 18+ formats, including DOCX, PDF, SRT and VTT. Check the export on the plan you'll buy, not on the top one.

  • Is it legal to transcribe a recorded conversation?

    Transcribing is legal when the recording itself is, and that depends on where you are: California, for example, bars recording a confidential communication without the consent of all parties. Meeting bots that join automatically can record people who never agreed, so announce the bot and turn off auto-join when outsiders are on the call.

  • How do I make AI transcripts more accurate?

    Fix the audio before the software: one microphone per speaker, no echo, no music behind voices. Then set the right language before uploading and add custom vocabulary for names, brands and jargon. Check how much your plan allows: Otter's free plan takes 5 vocabulary terms and Pro 100 names plus 100 other terms, while Notta lists custom vocabulary as a Pro feature, for English and Japanese only.