Human vs AI Transcription: Accuracy, Cost, Speed

On clear, single-speaker audio, AI transcription now comes close to human accuracy. OpenAI's gpt-transcribe model lists at under half a cent per audio minute ¹, and in one 2025 test Apple's speech API returned a transcript of a 7-minute-31-second episode in 9 seconds ². Human services still win on noisy, accented, overlapping or specialist recordings, and wherever a transcript has to be certified. For most people the best value is AI first, then a quick human check of the parts that matter.

The sections below show where each number comes from and what it does not prove. The main measure used is word error rate, the share of words a system gets wrong.

On clear audio AI is close to human accuracy; on hard audio humans still win

Clean, read English is where AI is strongest. Whisper Large V2 had a word error rate of 2.7 percent on the LibriSpeech Clean test set ³. The model's authors conclude that its English recognition is "not perfect but very close to human-level accuracy" ⁴.

Hard audio is a different test. The same model made 25.5 percent word errors on the CHiME6 test set ⁵, about 9.4 times its error rate on clean speech ⁶. Whispered speech was harder again: a 2024 baseline test measured 18.8 percent word error rate for Whisper on whispered speech ⁷.

Human transcribers are not a fixed yardstick either. In a 2017 Microsoft study, professional human transcribers made 5.9 percent word errors on the Switchboard telephone conversations and 11.3 percent on the harder CallHome conversations ⁸. Microsoft's automated system scored 5.8 percent and 11.0 percent on the same two sets ⁹, slightly better than the professional transcribers. An IBM study of the same kind of test reported a 6.8 percent word error rate for human transcription on CallHome ¹⁰. Different teams report different human scores, so one human figure is not a fixed standard.

The pattern holds across these tests. AI matches careful human work on clean or moderately hard speech, and falls behind as the audio gets quieter, further from the microphone, more overlapped or more specialist. Those studies date from 2017 to 2024 ^{8 7}, and newer models may do better, so read the table as a direction rather than a current ranking. For the wider picture, see how accurate speech recognition is. The short answer to the question is on the question page for AI transcription accuracy.

An accuracy claim means little until you know the audio and how errors were counted

Word error rate is the proportion of word errors to the words processed ¹¹. Insertions count as errors too, so the score can pass 100 percent ¹². Two numbers are comparable only when the audio, the scoring rules and the reference transcript match. Three things move a score more than the product name does.

• Distance from the microphone. In the Ravanelli figures from 2017, a single distant microphone gave a word error rate about 3.2 times the close-talking rate ¹³. The two conditions are not identical, so treat the ratio as rough.
• Scoring rules. The Open ASR Leaderboard scores normalized text, with punctuation and casing removed, and disfluencies such as filler words dropped ¹⁴. Your transcript may need those words, so a leaderboard score does not guarantee your transcript's quality.
• Sample size. A single recording says little about the next. In a 2025 9to5Mac test of one 7-minute-31-second episode, Apple's speech API had 10.3 percent word errors ¹⁵, and Whisper Large V3 Turbo had 1.5 percent ¹⁶. The author adds that the exact numbers would differ under slightly different methods ¹⁷.

Vendor claims need the same scrutiny. Rev's price page promises "99%+ guaranteed in 12 hours or less" for its human service ¹⁸, but the page does not say which audio the guarantee covers or how errors are counted. ElevenLabs says its Scribe v2 model has over 98 percent accuracy, yet its pricing page gives no test set or method for that figure ¹⁹. Treat both as claims to test on your own recordings.

AI also fails in a different way from people: it can invent text. The Whisper authors list complete hallucination, a transcript entirely unrelated to the actual audio, among its known failure modes ²⁰. A 2025 study found that when non-speech sounds were played to the model, a set of hallucinated phrases appeared frequently ²¹. A low average error rate can hide a few invented sentences, so check silences, music and overlapping talk in any AI transcript. For a plain-language guide to reading these scores, see word error rate explained.

AI costs fractions of a cent per minute; human transcription costs far more

AI transcription is billed per minute or per hour of audio, at fractions of a cent per minute. OpenAI's list price for gpt-transcribe is $0.0045 per minute ¹, and gpt-4o-mini-transcribe is $0.003 per minute ²², about $0.18 for one hour of audio ²³. Rev charges $1.99 per minute for human transcription ¹⁸, the one current human price we verified for this guide. That comes to about $119.40 an hour, and against gpt-transcribe the gap is about 442 times ²⁴.

The chart shows the gap, not a ranking of AI services. The prices behind it are listed below.

Several AI services also have free allowances. AWS Transcribe covers up to 60 audio minutes a month for the first 12 months after signup ²⁹, and Azure's free tier includes 5 audio hours a month of standard real-time transcription ³⁰. Those prices were checked in October 2026 and will change, so confirm them on the vendor's page before you commit. For a fuller comparison of cloud and on-device costs, see AI dictation cost. For a broader look at the options, see how to transcribe audio to text.

Price is not the whole cost. An AI draft with errors still needs review, and that review takes time. A cheap transcript that needs an hour of correction can cost more than a human service that delivers a finished file.

AI returns a transcript in minutes; human services take hours to days

AI engines run far faster than real time. The Open ASR Leaderboard measures speed with an inverse real-time factor ³¹. On its English short-form track, Parakeet TDT 0.6B v2 reached 3,390 and OpenAI's Whisper Large v3 reached 146 ^{32 33}. In the 9to5Mac test, Apple's speech API took 9 seconds to transcribe the 7-minute-31-second episode ², and Whisper Large V3 Turbo took 40 seconds, the longest of the three engines ³⁴.

Human services are measured in hours or days. Rev's human service guarantees 99 percent or better in 12 hours or less ¹⁸. Court transcripts can take longer still. A 2013 order from the District of Colorado set the maximum rate for an ordinary transcript, due within 30 days, at $3.65 per original page ³⁵. Human speed also shows in official tests. Delhi's 2022 English stenographer test allowed 8 minutes of transcription for each minute of dictation ³⁶, which is a test allowance rather than a service promise.

Turnaround here is the time for the transcript to arrive, not the time to a finished, checked document. If you want a deeper answer on what an hour of audio takes, see how long it takes to transcribe an hour of audio. The turnaround time of a service is the number to compare.

Legal, medical and live captioning work still calls for a human pass

For court records, the rule is about certification, not only accuracy. Under the federal statute on court reporters, a transcript certified by the reporter "shall be deemed prima facie a correct statement" of the testimony ³⁷. No transcript counts as official unless it is made from records certified by the reporter or another designated person ³⁸. This is the federal rule, and state rules differ. An AI transcript of a hearing, however accurate, is not an official record unless a designated person certifies it.

The professional bar is high. To pass each leg of the NCRA Registered Professional Reporter skills test, a candidate needs 95 percent accuracy ³⁹. NCRA's online skills test for the Certified Realtime Captioner requires 96 percent accuracy at 180 words per minute over five minutes ⁴⁰. The U.S. Bureau of Labor Statistics expects AI-generated transcription to dampen employment demand for court reporters and simultaneous captioners ⁴¹. For the job side, see transcription jobs and the legal transcription guide.

Live captions are the hardest case, because there is no time to edit. The Justice Department's ADA guidance describes real-time captioning, also called CART, as a service similar to court reporting in which a transcriber types what is said ⁴². A 2024 study reports that CART accuracy can decline considerably, which affects how well deaf and hard of hearing people follow the content ⁴³. It also finds that existing ways of improving automatic speech recognition still fail to match trained CART captioners ⁴⁴. For this need, see speech to text for deaf and hard of hearing people.

In medicine, the risk is invented or missed content. A 2025 study of LLM-generated clinical notes, built from 12,999 clinician-annotated sentences ⁴⁵, found a hallucination rate of 1.47 percent ⁴⁶ and an omission rate of 3.45 percent ⁴⁷. Those are AI-written notes rather than plain transcripts, so the numbers do not transfer directly, but they show why clinical text needs a human check. See how AI is changing medical transcription and AI medical scribes.

AI first, then human review, is usually the best value

Use AI for the first draft, then review it. Microsoft's guidance for dictating messages to Copilot tells users to review the transcribed content for accuracy, because AI may occasionally misinterpret speech ⁴⁸. The table below sets out which route fits which kind of recording.

A practical workflow for a recording you need in text:

1. Record close to the microphone. Distance raises errors ¹³, so a headset or a lapel microphone beats a laptop across the room.
2. Run a first pass with AI. On a Mac, Apple says Voice Memos can recognize and transcribe speech in recordings ⁴⁹, though this is not available in every country or region ⁵⁰. Apple says the feature works with macOS Sequoia on a Mac with Apple silicon ⁵¹. For a finished file in the browser, the free audio to text tool is another option. For other free choices, see how to transcribe audio to text.
3. Check the risky parts. Look at names, numbers, silences, music and overlapping speakers, since those are where errors and invented text cluster.
4. Send only what must be perfect to a person. Pay a human service for the hard sections, or use a designated reporter when the transcript is a legal record. If you need verbatim detail such as filler words, see verbatim vs clean verbatim.

Without AI, the choices are a human service, a stenographer, or your own typing from the recording. Typing speed sets how long the last route takes, which is covered in typing speed. Yap, which we make, is an on-device dictation app for Mac and Windows, so it suits writing you speak yourself rather than transcribing a finished file. Current prices are on the pricing page.

Whether transcription software is generative AI depends on the definition

Whisper, one widely used engine, is built on an encoder-decoder Transformer. The authors chose that architecture because they judged it well validated to scale reliably ⁵². Whether that counts as generative AI depends on how you define the term, and the answer is a matter of definition, not a fact about the audio. For the full discussion, see is speech to text AI? and the answer on generative AI.

For more on how the pieces fit together, the guides on meeting transcription, interview transcription and Whisper, ChatGPT and Gemini for transcription cover the common cases in depth. The short answers to the questions people ask are on how much transcription costs per minute, and how to get a transcript without AI. For the whole subject, see our overview of transcription.