Should a company self-host Whisper? License, GPU cost and where the audio goes (October 2026)

How I put this together

I read the README, license file and model cards of the four projects, then the pricing pages of five vendors, and compared OpenAI's data-controls table with Article 28 of the GDPR. Prices per minute are multiplied by 60 so that everything is per audio hour. I did not create an account, run a model or upload audio.

Four things did not line up. I logged a correction for our Speechmatics entry, from $0.129 to the $0.12 per hour the vendor shows today. I found that the FunASR README names no lab, so I logged a correction for our line that credits one. The Hugging Face cards for Whisper large-v3 and turbo give different licenses than the GitHub README, which I report below. And I left out a European GPU host: Hetzner's page loads its prices by script, so I took RunPod's published list prices instead and say so where it matters. How we review explains the general method.

Sources: [1][2][3][4]

Step 1: does the audio have to stay in your network?

This question comes first because its answer removes options. Article 28(3) of the GDPR says processing by a processor must be governed by a binding contract that sets out subject matter, duration, type of data and the controller's rights. Any API that receives recordings of identifiable people is such a processor. A model on your own server has no outside processor for the transcription itself.

I read what the largest API vendor documents. OpenAI's data-controls page says data sent to the API is not used to train its models unless you opt in. For the transcription endpoint, its table lists abuse-monitoring retention as none and marks the endpoint eligible for zero data retention. The European endpoint lists audio transcription too, with a note that it requires modified abuse monitoring or zero data retention, and data residency controls need a conversation with sales. For models released on or after March 5, 2026, residency endpoints carry a 10 percent uplift.

What this means for you: the common claim that an API keeps your audio for 30 days does not match OpenAI's own table for this endpoint, where the 30 days apply to others such as speech output. Self-hosting does not buy shorter retention there. It buys the absence of any third party processing the audio, which matters when your contracts, regulator or customers exclude processors. I read only OpenAI's page, so check the vendor you would actually choose.

Sources: [5][6]

Step 2: which license covers what you download

Whisper's README says the code and the model weights are released under the MIT License. GitHub lists WhisperX as BSD-2-Clause and pyannote.audio as MIT. For a company, those three lines look settled.

The speaker-labeling model is where it changes. pyannote.audio's current open pipeline, community-1, carries CC BY 4.0 according to its Hugging Face metadata, and you must accept user conditions and create an access token before you can download it. WhisperX's README sends you to the same step for diarization. Creative Commons states that CC BY 4.0 allows commercial use if you give credit, link the license and mark changes. The older pyannote 3.1 pipeline is MIT but gated in the same way. Read the conditions on the model page as a legal team would; I did not accept them, because that needs an account. pyannote.audio is the project behind both.

FunASR splits it more sharply. The toolkit is MIT, and its README says model weights are licensed separately, card by card. I opened three cards. SenseVoiceSmall points to the FunASR Model Open Source License Agreement, version 1.1, which requires attribution and keeping model names. That agreement also says users who engage in unjustified denigration or malicious smearing of the software automatically lose all licenses, and that the models are provided for reference and learning purposes only. The cards for Fun-ASR-MLT-Nano and paraformer-zh say Apache 2.0.

Sources also disagree on Whisper itself. The GitHub README says MIT for the weights, while Hugging Face lists Apache 2.0 on the large-v3 card and MIT on large-v3-turbo. I did not resolve this. Pull weights from one named location and record the license you saw there.

What this means for you: count licenses per pipeline, not per project. Whisper plus pyannote community-1 is MIT, CC BY 4.0 and a Hugging Face condition. Ask legal to list them before a pilot, and read the model card of every FunASR checkpoint you plan to ship.

Sources: [7][8][9][10][11][2][12][13][14][15][16]

Step 3: pick the stack that matches the job

Whisper alone reads a whole file and works through it in sliding 30-second windows. Its model card says it cannot be used for real-time transcription out of the box, and it returns no speaker labels. The README lists six sizes, from about 1 GB of video memory for tiny to about 10 GB for large, and a turbo model at about 6 GB that runs roughly 8 times faster than large on an A100 with English speech.

WhisperX wraps Whisper for recorded calls. Its README claims 70 times real time with large-v2 and less than 8 GB of GPU memory, and adds word timestamps and speaker labels through pyannote. The faster-whisper project, which WhisperX builds on, publishes a smaller benchmark: 13 minutes of audio in 1 minute 3 seconds in fp16 on an RTX 3070 Ti, and 16 seconds with batching in int8. Both are the maintainers' own numbers, and I did not reproduce them.

FunASR matters if you need live captions on your own servers. Its README lists a streaming Paraformer checkpoint for chunked live transcription, but that checkpoint covers Chinese and English. Fun-ASR-MLT-Nano covers 31 languages. For German or French calls, I would shortlist the Whisper family. I did not check German error rates; Whisper's README points to per-language figures in its paper.

What this means for you: for recorded client calls with speaker labels, the realistic stack is Whisper or faster-whisper, WhisperX and pyannote, run as a nightly batch. Live captions during the call are a different project.

Sources: [7][17][18][19][2]

Step 4: a month of audio costs less than the server that could transcribe it

Hosted list prices, converted to dollars per audio hour on October 10, 2026: OpenAI gpt-4o-mini-transcribe $0.18 ($0.003 per minute), gpt-transcribe $0.27 ($0.0045), Deepgram Nova-3 monolingual pay as you go $0.258 ($0.0043), Speechmatics Melia 1 $0.12 and Gladia Starter $0.61 for asynchronous files. OpenAI's pricing page I read lists no row for whisper-1. Our directory entries for Deepgram and Gladia match the vendors' pages.

For hardware I took RunPod's pricing page, updated September 27, 2026: a 24 GB L4 at $0.59 per hour and an RTX 4090 at $0.89. Over a 730-hour month that is $430.70 and $649.70 for a GPU that is always on. These are not European prices, and the page does not tie a rate to a region, so treat them as an anchor for the order of magnitude.

The break-even sits far above a mid-sized company's volume. The always-on L4 equals the API bill at 706 audio hours a month against Gladia, 1,595 against gpt-transcribe and 3,589 against Speechmatics. For 300 hours, the insurance broker pays $36 to $183 a month to an API and $430.70 for the GPU, before anyone maintains it. Deepgram also sells Whisper Large as a hosted model at $0.0048 per minute, $86.40 for 300 hours, so Whisper without operations exists.

There is a best case. At 13 minutes in 63 seconds, a speed of about 12.4 times real time from the faster-whisper README, one audio hour needs about 5 minutes of GPU. At $0.59 an hour that is about $0.048 per audio hour, about 40 percent of the cheapest API. That benchmark ran on a different card, and it assumes the GPU never idles. An idle hour is billed like a busy one.

What this means for you: the deciding numbers are utilization and staff time, not the price per hour. No source publishes the cost of an engineer to patch drivers, watch failed jobs and update models. The calculator leaves it out; add your own hours on top.

Sources: [20][21][1][22][4][19]

Step 5: what can go wrong after go-live

A 2024 study of OpenAI's Whisper service in 2023 found that roughly 1 percent of transcriptions contained entire phrases or sentences that were not in the audio, and 38 percent of those hallucinations included explicit harms such as perpetuating violence or implying false authority. I ran no measurement on current models, so treat the figure as a reason to plan review, not as today's error rate.

Whisper's own model card recommends against use in high-risk domains such as decision-making contexts, and cautions against transcribing recordings of people taken without consent. A transcript that becomes a record, a complaint file or an evidence trail needs a human check, whichever route you pick.

Upkeep differs by project. Whisper's latest GitHub release is v20250625, with the last code change on August 31, 2026. WhisperX released v3.8.6 on May 25, 2026, pyannote.audio 4.0.7 on June 30, 2026 and FunASR v1.4.16 on September 18, 2026. WhisperX asks for a CUDA toolkit 12.8 install before it. What this means for you: name an owner and a re-check date, because every one of these is a Python package you upgrade yourself.

Sources: [23][17][24][18][9][25]

Which route fits, and for whom it does not

Three routes came out of this: a hosted API, a hosted Whisper model, or your own server. Our directory lists 17 transcription tools and 4 open-source projects in the audio repo niche. The helper below ends on a starting point.

From the numbers I take this rule. Below about 700 audio hours a month, buy it; between that and roughly 3,600, the price decides little and data control decides more; above that, self-hosting pays only if someone owns the servers.

Sources: [20][4][26]