Whisper 自行部署與託管服務的成本打平點約為 85–250 小時音訊

Short answer: self-hosting Whisper costs nothing per minute and roughly 2–6 hours of setup plus ongoing maintenance. A metered hosted service costs around €0.02/minute and nothing to set up. At those rates you'd need to process roughly 85–250 hours of audio before the setup time pays for itself at any reasonable hourly rate — and that calculation ignores maintenance entirely. Below that, hosted wins on economics. Self-hosting wins outright on one axis that has nothing to do with cost: the audio never leaves your machine.

Whisper being open source and free gets stated a lot and is true in the sense that the weights cost nothing. The confusion is between free of licence fees and free, and the gap between those two is where people lose weekends.

What does self-hosting Whisper actually involve?

Self-hosting Whisper has three viable paths, in ascending order of effort:

whisper.cpp — a C/C++ port with no Python dependency. Compiles in a couple of minutes, runs on CPU, uses Metal on Apple Silicon and CUDA on Nvidia. This is the path of least resistance and the one to try first.

git clone https://github.com/ggerganov/whisper.cpp
cd whisper.cpp && make
sh ./models/download-ggml-model.sh large-v3
./main -m models/ggml-large-v3.bin -f audio.wav -osrt

That's genuinely most of it, plus converting your input to 16 kHz WAV with ffmpeg first, which is the step that trips people up.

faster-whisper — a CTranslate2 reimplementation, several times faster than the reference implementation with lower memory use. Python, so you're managing an environment, and CUDA if you want GPU speed, so you're managing driver versions. Meaningfully faster in exchange for meaningfully more setup.

Reference openai/whisper — the original PyTorch implementation. Slowest of the three and the heaviest dependency tree. Use it if you need something the ports don't implement; otherwise don't.

How fast is it on hardware you already own?

Throughput is where expectations break. Rough figures for large-v3, which is the model you want if accuracy matters:

These vary widely with audio length, quantisation, batch settings and thermal behaviour — treat them as orders of magnitude, not specifications, and benchmark your own hardware before planning around any of them.

The practical consequence: on a laptop without a discrete GPU, large-v3 runs close to realtime. Transcribing a one-hour meeting means an hour of a hot, loud, largely unusable machine. That's fine for occasional overnight batches and unworkable for "I need this transcript before my next call."

Many people respond by dropping to base or small, which is fast enough — and at that point you're comparing a small model on your laptop against large-v3 on someone's server, which is not the comparison you started out making.

What does the cost comparison look like honestly?

Assume €0.02/minute metered (€1.20/hour of audio) and value your setup time at €50/hour.

So: below roughly 100 hours of audio processed, hosted is cheaper once your time is priced. Above a few hundred hours, self-hosting clearly wins, and above a few thousand it isn't close.

Cloud GPU changes this unfavourably. Renting a GPU at €0.50–1.50/hour to transcribe at 20× realtime works out to roughly €0.03–0.08 per hour of audio in compute — cheaper than metered per unit, but you're now paying for instance time including idle, plus setup, plus orchestration. It only makes sense as a batch job over a large archive.

So when should you self-host?

Four situations where it's the right answer regardless of the arithmetic:

The audio can't leave your infrastructure. Legal material under a protective order, clinical recordings without a signed BAA, anything under a contract barring third-party processing. This is the strongest case and it's not about money. No vendor policy substitutes for the file never leaving the building.

You're processing an archive. Ten thousand hours of back-catalogue is €12,000 metered and a weekend of GPU time self-hosted. Obvious.

You need to modify the pipeline. Custom VAD, a specific diarisation model, forced alignment, domain-adapted decoding, integration into an existing ML stack. Hosted tools give you their pipeline, not yours.

Air-gapped or unreliable connectivity. Field research, secure facilities, transcription on a plane.

And where it isn't:

You need it now, occasionally. The setup cost never amortises across four files a month.

You want the surrounding features. Diarisation that works out of the box, summaries, subtitle export, translation, a UI other people can use. Whisper produces text. Everything else in a transcription product is the other 80% of the work, and rebuilding it is a project, not an afternoon.

You're not the person who'll maintain it. A Whisper install that only one person on the team can fix is a liability disguised as a saving.

What about accuracy — is a hosted tool better?

Usually not, and this is worth being clear about because it cuts against the vendor's interest.

Most hosted transcription tools run the same open models you'd run yourself, or commercial models in the same accuracy band. Published benchmarks put Whisper large-v3 around 2–3% WER on clean read English, rising to roughly 8% on mixed real-world audio and higher on poor-quality call recordings. Competing commercial engines cluster in the same range, trading places depending on the test set — AssemblyAI and Deepgram lead on some, Whisper on others, particularly outside English where its multilingual coverage is unusually broad.

The honest summary: on the same model, self-hosted output is the same as hosted output. What hosted services add is preprocessing, VAD, chunking strategy and post-processing, which do move real-world accuracy — but they move it by a few points, not by a category.

If you self-host, the biggest accuracy lever available to you is not the model. It's the input: 16 kHz mono, one speaker per channel where possible, and initial-prompt hints for domain vocabulary.

A reasonable decision procedure

1. Estimate your annual hours of audio. Under 50, use a hosted tool and stop.
2. Check whether the data can legally leave your control. If not, self-host regardless of volume.
3. Check what hardware you have. Apple Silicon or a discrete Nvidia GPU makes self-hosting pleasant. Neither makes it painful.
4. Try whisper.cpp on one real file before committing either way. Thirty minutes, and it tells you more than any comparison table.
5. Price the features you'd have to rebuild — diarisation, subtitles, summaries, a UI — before treating "free" as free.

Frequently asked questions

Is Whisper free to use?
The model weights are free and openly licensed. Running it costs hardware, electricity and your setup and maintenance time. For low volume, those exceed the per-minute cost of a metered hosted service.

What hardware do I need to run Whisper large-v3?
Around 4–6 GB of RAM or VRAM for a quantised build. It runs on CPU at roughly realtime speed, which is usable for batch jobs and frustrating for interactive use. Apple Silicon with Metal or a consumer Nvidia GPU brings it to several times realtime.

Is whisper.cpp as accurate as the original Whisper?
Substantially, yes, for the same model size. Aggressive quantisation costs a small amount of accuracy in exchange for speed and memory. The model you choose (tiny through large-v3) affects accuracy far more than the implementation does.

Does a paid transcription service use a better model than Whisper?
Often it is Whisper, or a model in the same accuracy band. Published WER benchmarks put the leading commercial engines and Whisper large-v3 within a few points of each other, with the ordering depending on the test set. Paid services differentiate on pipeline engineering, features and reliability rather than raw model quality.

Disclosure: I work on convert.express, a hosted Whisper, MAI, Gemini transcription service. If your answer after reading this is "I'll self-host," that's a correct outcome — the break-even is real and it isn't always on our side.

Sources for the benchmark figures: published WER comparisons of Whisper large-v3, AssemblyAI Universal-2 and Deepgram Nova-3 (2026). Benchmark results vary by test set; run your own audio through candidates before deciding.