Compare · Transcription
Whisper vs. cloud transcription: Deepgram, AssemblyAI and Rev
Run Whisper on your own computer or pay a cloud service? Cost per hour, privacy, published accuracy claims and the hallucination caveat (November 2026).
We may earn a commission if you buy through links on this page. It never changes what we recommend. How this works
If your audio is sensitive and you have a decent computer, run Whisper locally: it costs nothing per hour and the audio never leaves your machine. If you want less setup, speaker labels and a support team, pay a cloud service. The per-hour prices are low enough that cost rarely decides it. Privacy, setup effort and accuracy checks do.
This is a research-based comparison: we read vendor pricing and documentation, the Whisper paper and news coverage (sources listed with this article). We haven't run our own accuracy tests, and we don't quote accuracy numbers that we can't source. All prices are as of November 14, 2026 and change.
At a glance
| Whisper, local | OpenAI hosted API | Deepgram | AssemblyAI | Rev | |
|---|---|---|---|---|---|
| Cost | $0 plus your hardware | whisper-1 $0.006 per min; gpt-transcribe $0.0045 per min; gpt-4o-mini-transcribe about $0.003 per min | Nova-3 pre-recorded: $0.0043 per min monolingual, $0.0052 multilingual (pay as you go) | Universal-3.5 Pro $0.21 per hour; Universal-2 $0.15 per hour | Free 45 AI min; AI plans from $25.49 a seat a month billed annually; human transcription from $1.99 per min |
| Free credit | n/a | Not listed on the pricing page we read | $200, no expiry, no card | $50, no card | Free plan, 45 AI min |
| Speaker labels | Not built in | A separate diarize model at about $0.006 per min (estimate) | Check Deepgram's docs | Diarization +$0.02 per hour | Not confirmed |
| Where audio goes | Stays on your computer | OpenAI | Deepgram | AssemblyAI | Rev |
| Setup | You install and run it | API call | API call | API call | Upload on the web or app |
OpenAI lists some model prices as estimates because they are billed by tokens. Deepgram's per-hour figure is our arithmetic ($0.0043 per minute is about $0.26 per hour).
Cost on a real workload
Say you record 10 hours a month.
- Whisper on your own computer: $0 in fees.
- OpenAI whisper-1 at $0.006 a minute: $3.60.
- OpenAI gpt-transcribe at $0.0045 a minute: $2.70.
- Deepgram Nova-3 monolingual at $0.0043 a minute: about $2.58.
- AssemblyAI Universal-3.5 Pro at $0.21 an hour: $2.10, or $2.30 with standard diarization (+$0.02 an hour).
- Rev human transcription at $1.99 a minute: $1,194.
That last line is why the choice isn't really about machines being cheap. Machine transcription costs a few dollars for 10 hours. Human transcription costs hundreds. Rev's page says its human transcripts are 99% or more accurate and delivered in 12 hours or less. That is Rev's own claim and we haven't checked it. If you need a transcript for court or a published quote, a human review is the thing to pay for.
Whisper, run locally
OpenAI released Whisper's code and model weights under the MIT License. Its GitHub page lists six model sizes. The smallest (tiny, 39 million parameters) needs about 1 GB of video memory and is the fastest. The largest (large, 1.55 billion parameters) needs about 10 GB. A "turbo" model, an optimized version of large-v3, needs about 6 GB and runs about eight times faster than large, with minimal accuracy loss, per the page. Community ports such as whisper.cpp run on a regular computer processor, including Apple Silicon, with no network connection needed.
Accuracy, as published. The Whisper paper (arXiv, 2022) says the models were trained on 680,000 hours of multilingual audio and that, compared with humans, they "approach their accuracy and robustness." The README says performance varies widely by language and the English-only models tend to do better. We don't know how that translates to your meeting audio.
Hallucinations. A caveat from news reporting: in October 2024, the Associated Press reported that researchers found Whisper sometimes invents text. A Cornell and University of Virginia study found 187 hallucinations in over 13,000 clear audio snippets, and nearly 40% were harmful or concerning. A University of Michigan researcher reported hallucinations in 8 of every 10 transcripts he checked. These findings date from 2024 and cover versions then in use. OpenAI's own disclosures warn against using Whisper in high-stakes decisions. It doesn't mean that every transcript is wrong. It means you shouldn't quote someone from a transcript without hearing the audio.
Best for: privacy-minded users, people with lots of audio, and anyone comfortable with a terminal or a ready-made app.
Downsides:
- You do the setup, updates and troubleshooting.
- No speaker labels out of the box.
- Long files on a slow computer take a long time.
- Large models need a strong graphics card or a recent Mac.
- Invented text is possible. Check important passages.
OpenAI's hosted API
OpenAI's pricing page lists whisper-1 at $0.006 a minute, gpt-transcribe at $0.0045 a minute and gpt-4o-transcribe at an estimated $0.006 a minute, with a gpt-4o-mini-transcribe tier at an estimated $0.003. It also lists a diarization model. Third-party sites mention retirement dates for older models; we couldn't confirm those on OpenAI's page, so check before building on them.
Downsides: your audio goes to OpenAI; per-minute billing; model names and prices shift, and we couldn't confirm what happens to older ones. The hallucination caveat above applies to Whisper-based models. We haven't confirmed it for the newer ones.
Deepgram
Pay-as-you-go Nova-3 pre-recorded is $0.0043 a minute for one language and $0.0052 for multilingual, billed per second. New accounts get a $200 credit with no expiry. Deepgram's model improvement program trains on submitted data unless you opt out with a parameter (mip_opt_out=true); community answers say opting out affects your rates, so ask before you rely on it.
Downsides: developer-oriented, and privacy is a setting you must set. Check for yourself how data is retained.
AssemblyAI
Universal-3.5 Pro is $0.21 an hour (18 languages), Universal-2 is $0.15 an hour, and add-ons stack on top: standard diarization is +$0.02 an hour, keyterms prompting +$0.05, Medical Mode +$0.15. Multichannel audio is billed per channel. New accounts get $50. Its documentation says paid plans can opt out of model training from the dashboard, while free accounts can't opt out, set retention or sign a BAA.
Downsides: developer-oriented; the free account has fewer privacy controls than paid ones; add-ons raise the price.
Rev
Rev sells an app and plans: Free (45 AI minutes), Essentials ($25.49 a seat a month billed annually, $29.99 monthly, 5,000 AI minutes), Pro ($47.99 / $59.99, 10,000 minutes) and a custom Unlimited plan. Human transcription starts at $1.99 a minute, and legal transcripts are sold per page.
Downsides: much costlier than raw APIs if you use human transcription, and seat plans add up for teams. Its accuracy claims are its own.
Privacy, in short
Local Whisper is the only option here where the audio doesn't leave your computer. Cloud services process your audio on their servers, and their training and retention terms differ by vendor and plan. If someone else's voice is in the file, they may have a say too. See recording consent. Not legal advice.
Who should pick which
- Local Whisper: sensitive audio, a lot of it, and some patience for setup. Check the output.
- A cloud API (Deepgram, AssemblyAI or OpenAI): you're building something, or want speaker labels and speed, and the audio is okay to send. Opt out of training if you can.
- Rev (human): the transcript will be quoted, filed or used in a legal or published setting.
- Neither: if you just want your recordings turned into readable pages, an app that does it for you is easier. Many recorders and apps include transcription. See best voice recorders and apps.
Money and independence
We did not confirm an affiliate program for any of these services. Our verdicts don't depend on one. Links: Whisper, OpenAI API, Deepgram, AssemblyAI, Rev.
Daytapes is our own product (invite-only at launch). You can attach Whisper exports (.txt, .srt, .vtt, .json) along with the audio, so a local transcript and your recording end up on one page. See Daytapes as your recorder.
Questions
Is Whisper free?
The code and model weights are released under the MIT License, so running it yourself costs nothing but your hardware and time. OpenAI's hosted transcription API is a paid service (as of November 14, 2026).
How much does cloud transcription cost per hour?
Listed pay-as-you-go rates: about $0.15 to $0.21 per hour at AssemblyAI, about $0.26 per hour for Deepgram Nova-3 monolingual, and about $0.18 to $0.36 per hour (our arithmetic) for OpenAI's transcription models. Rev charges from $1.99 per minute for human transcription.
Can Whisper make things up?
Yes. A 2024 Associated Press report cited researchers who found invented text in Whisper transcripts. Check important quotes against the audio.
Does cloud transcription train on my audio?
It depends on the vendor and plan. Deepgram and AssemblyAI both have model-improvement programs with opt-outs. Read the terms before uploading anything sensitive.
Sources
- OpenAI: Whisper on GitHub (license, models, VRAM) (accessed November 14, 2026)
- Radford et al.: Robust Speech Recognition via Large-Scale Weak Supervision (arXiv 2212.04356) (accessed November 14, 2026)
- OpenAI API pricing (accessed November 14, 2026)
- Deepgram pricing (accessed November 14, 2026)
- Deepgram: Model Improvement Partnership Program (accessed November 14, 2026)
- AssemblyAI pricing (accessed November 14, 2026)
- AssemblyAI: Data retention and model training (accessed November 14, 2026)
- Rev pricing (accessed November 14, 2026)
- AP via Anchorage Daily News: Researchers say an AI-powered transcription tool used in hospitals invents things no one ever said (October 26, 2024) (accessed November 14, 2026)
- Linuxlinks: whisper.cpp (accessed November 14, 2026)