Home → Blog → Self-hosted Whisper or a transcription API: what costs less
Self-hosted Whisper or a transcription API: what costs less
Whisper is open, a few gigabytes in size and starts with one command, so the urge to run it yourself is understandable. But hardware bills arrive monthly while audio arrives by the minute, and at small volumes self-hosting loses to an API by a wide margin. Let us do the arithmetic.
What self-hosting actually includes
Loading the model is the easy part. The rest shows up a month later:
- A GPU. large-v3 needs about 10 GB of VRAM, so a 12 GB card or better.
- A queue. Two concurrent requests do not fit on one card — you need a scheduler.
- Format conversion. Clients send m4a, ogg, webm, so you maintain ffmpeg and chase codecs.
- Someone on call. Drivers, CUDA upgrades, restarts after a crash — that is a person's time.
Running the numbers
Take a typical load: 500 minutes of audio a month.
| Option | Monthly spend | Also needed |
|---|---|---|
| Rented cloud GPU | from $100 for a modest card | setup, upgrades, monitoring |
| Your own card | $400–800 up front plus power | space, cooling, downtime on failure |
| Pay-per-minute API | 500 minutes at the per-minute rate | nothing |
A rented card bills around the clock, including the hours the queue is empty. That is where the "self-hosting is cheaper" estimate falls apart: 500 minutes of audio is under nine hours of GPU work out of 720 available.
When self-hosting wins
Two cases justify your own server.
Sustained high volume. Once the card is busy at least half of every day, its cost spreads across thousands of minutes and the per-minute price drops below any API.
Data that must not leave the perimeter. Healthcare, banking, classified work — there the answer is not about price. Everywhere else it is worth checking how the provider handles recordings: if they are deleted within a day and never used for training, the risk is comparable to any cloud service you already use.
A sane order of operations
Starting with self-hosting is almost always premature: you know neither the accuracy on your own audio nor the real volume. The working sequence is:
- Connect to an API and run real recordings through it — that is a day of work.
- Collect a month of minute counts and look at the actual invoice.
- If the volume grew enough to keep a card busy, move it in-house and keep the API as overflow capacity.
Because the protocol is the same in both directions, migrating means changing a base URL and a key — the code does not move at all.
Try it on your own recordings. Sign-up takes a minute, and the free minutes are enough to judge the quality.
Get a free API keyFrequently asked questions
At what volume does a GPU pay for itself?
Roughly when it stays busy half the day. At a few hundred minutes a month the card idles and per-minute billing is cheaper.
Can I run Whisper without a GPU?
Yes, on CPU, but tens of times slower: a minute of audio takes minutes. Fine for occasional files, unworkable as a pipeline.
How hard is it to move in-house later?
The protocol is identical, so the client code changes a URL and a key. Everything else — queue, conversion, monitoring — you build from scratch.
What about privacy when using an API?
Look at retention. Here recordings and transcripts are deleted within 24 hours and never used to train models.
Related reading
- Which Whisper model to choose for speech recognition — How Whisper models differ from tiny to large-v3: speed, accuracy, VRAM requirements and when a smaller model is genuinely the better trade.
- Speech-to-text pricing: what you actually pay for — How transcription billing works, which parts of your audio cost money, when self-hosting a GPU is cheaper than an API, and how to estimate your own monthly spend.
- Transcription privacy and security: what to check with a provider — Five questions to ask a speech recognition provider: retention, training on your data, channel encryption, staff access and deletion on request.