If you transcribe interviews - as a journalist, a qualitative researcher, a podcaster, or a PhD student with fifty hours of fieldwork - the decision is mostly arithmetic. Prices in this market span three orders of magnitude for the same audio, and the right answer depends almost entirely on how many hours you have and how accurate the result needs to be.

Short answer: run MacWhisper, Buzz, or Vibe locally for free unless you have a specific reason not to. If you need speaker labels and a polished web editor, Sonix at $10 an hour is the best paid AI option. If the transcript has to be right - court, publication, poor audio - pay Rev $1.99 a minute for a human.

What it costs, per hour of audio

The gap between $0 and $120 an hour for the same recording is the entire story of this market.

Free and local: the default answer

Whisper runs on a laptop. For file transcription - where you don't need real-time output - a local model on a modern machine transcribes an hour of audio in a fraction of that time, at no cost, with the recording never leaving your possession.

MacWhisper is the Mac standard: free for the tiny and base models, €59 one-time for Pro, with watch folders, YouTube URL ingestion, system audio capture, and subtitle export formats. Note that the App Store build is a separate SKU called Whisper Transcription at $29.99 a year or $99.99 lifetime.

Buzz is the cross-platform equivalent - macOS, Windows, Linux, free under MIT, around 20,900 GitHub stars - handling audio files, video files, YouTube links, and live microphone input with CUDA, Apple Silicon, and Vulkan acceleration.

Vibe is the one to pick if you need speaker diarization, which for interviews you usually do. Free, MIT, macOS, Windows, and Linux, with batch processing and exports to SRT, VTT, TXT, HTML, PDF, JSON, and DOCX. It supports Whisper, Nemotron 3.5, and Parakeet TDT v3.

For a researcher with confidential interviews, the privacy argument is at least as strong as the cost one: no upload, no third-party retention, no consent question about a processor your participants never agreed to. See cloud vs on-device dictation.

The trade-off is that you handle the workflow yourself - no web editor with synchronised playback, no collaboration, no sharing links.

Best paid AI: Sonix

Sonix prices honestly by the hour rather than by seat, which suits irregular workloads. Standard is $10 an hour pay-as-you-go with no subscription, single user. Premium is $5 an hour plus $22 per user per month, adding team collaboration and AI analysis. Break-even is around 4.4 hours a month per user - below that, stay on Standard.

Billing is prorated to the second, which matters when your files are 43 minutes rather than a round hour. 54+ languages, and a claimed 99% accuracy that, like every vendor figure in this industry, is measured on clean audio.

New accounts get 30 free minutes.

Cheapest AI with no subscription: Temi

Temi is Rev's budget service: $0.25 per audio minute, $15 an hour, pay-as-you-go with no subscription. You fund an account and draw against it. First file gets 45 free minutes with no card required. Web plus an Android app.

It's fully automated with no human review, so treat the output as a first draft. For a quick searchable transcript of a clean recording it's fine, and the absence of a subscription is worth something if you transcribe three times a year.

When to pay for a human

Automated transcription fails predictably: overlapping speech, strong accents, poor room acoustics, specialist terminology, and low-quality phone recordings. If your recording has two of those, budget for a human.

Rev charges $1.99 per audio minute for human transcription, with some tiers reported from $1.50 - roughly $119 an hour. It also sells AI transcription at $0.25 a minute, a free tier at 45 minutes a month, and subscriptions from $25.49 a seat per month annually. Rev Max starts at $29.99 a month for 20 hours of AI transcription.

Happy Scribe charges $2 a minute for human transcription with a claimed 99% accuracy, and is the stronger choice if you also need subtitles and captions. Its AI tiers run $8.50 a month annually for Basic with 120 minutes, up to $59 a month annually for Business with 6,000 minutes, with extra AI minutes at $0.20.

The maths that usually decides it: one hour of human transcription costs about what twelve hours of AI transcription costs, and about what a $59 lifetime licence for local software costs. If you're publishing quotes verbatim, that's cheap. If you're coding themes across forty interviews, it isn't.

For newsrooms: Trint

Trint is built for editorial workflows and priced for organisations. Starter is around $80 a seat per month, or $52 billed annually - and gets you seven files a month, which is a cap worth noticing. Advanced is around $100 a seat per month, or $60 annually, unlimited files, single user. Enterprise adds SSO/SAML, SCIM, audit logs, custom retention, and a full API. No free tier, and annual billing is required on most tiers.

Trint Live handles mobile capture. For a working journalist filing daily, the unlimited Advanced tier at $60 a month annually is roughly six hours of Sonix - so it wins above that volume, and loses below it.

For podcasters: Descript

If the transcript is a step towards editing audio or video rather than the final product, Descript is a different proposition: delete a word from the transcript and it vanishes from the recording. Free covers about 60 media minutes a month; Hobbyist is $16 a user per month annually, Creator $24, Business $50 with around 30 hours a user per month.

Getting better results from any of them

Transcription quality is set by the recording, not the software. In rough order of impact:

1. Use a real microphone. A $60 USB mic beats a laptop's built-in array by more than any software choice will. See the best microphone for dictation.
2. Record each speaker separately if you can. Overlapping speech is the single biggest cause of diarization failure.
3. Kill the room. Soft furnishings, closed windows, no fan. Reverb hurts more than background noise.
4. Record uncompressed. WAV or a high-bitrate M4A, not a compressed voice memo.
5. Say names and jargon clearly at the start. Some tools accept a custom vocabulary list; feed it every proper noun in the interview.
6. Budget editing time. Even at a claimed 99%, that's roughly one error every two lines.

Frequently asked questions

What's the cheapest way to transcribe interviews?

Free, using Buzz, Vibe, or MacWhisper locally on your own machine. The only cost is your time and electricity, and the audio never leaves your computer.

What's the most accurate transcription service?

Human transcription - Rev at $1.99 a minute or Happy Scribe at $2 a minute, both claiming 99%. Among AI services, quality differences are small compared with the difference recording quality makes.

Which transcription tool identifies speakers?

Vibe does it free and locally. Among paid services, Sonix, Rev, Trint, Happy Scribe, and Descript all offer speaker diarization.

Is AI transcription good enough for research?

For thematic analysis and searching, generally yes. For publishing verbatim quotes, check every quote against the audio, or pay for human transcription on the sections you'll quote.

How long does transcription take?

Cloud AI services take minutes. Local transcription depends on your hardware and model - an hour of audio might take five minutes on an M-series Mac with a medium model, or considerably longer on an old CPU with a large one. Human transcription takes hours to a day.

Last verified: 10 August 2026.