Transcription is priced by the duration of the audio, and the rate you are quoted moves with four things: whether a human touches the file, how quickly you need it back, the language and the recording conditions, and how many hours you commit to up front. A human-corrected transcript costs considerably more per hour than a machine-only one, and that gap is the largest single factor in any quote. The per-hour figure is the easy half of this question, though. If you are pricing an archive rather than one file, a different number decides your budget.
The routes, and what each one really charges you
Machine transcription through an API is billed by duration and sits at the bottom of the range. Quality tracks the audio: crosstalk, heavy accents, room noise, and industry jargon all push the error rate up, and the price does not go down to compensate.
Machine transcription plus in-house cleanup moves money from the invoice to your payroll. Someone reads along and fixes names, terminology, and speaker labels. Cleanup time scales with the error rate, so bad audio gets billed twice, once by the vendor and once by whoever has to sit through it.
Professional human transcription is the top of the range, and it is the right answer when the transcript is the deliverable. Published captions, legal records, and translation source material all have to be correct in a way that no automated pass guarantees.
Self-hosted open models produce no per-hour invoice at all. You pay in compute, in setup, and in one person owning the thing for as long as it runs. For a large back catalog that trade often works. For twelve hours a month it rarely does, because the fixed cost of ownership never goes away.
The last route is the one people forget they are on: transcription bundled into an editing or asset tool they already pay for. The per-hour price is invisible, and the constraint shows up as a quota rather than a bill.
The number that actually sets the budget
Split your hours into two piles. The back catalog is a one-time multiplication, and you can get it within a few percent by sampling a handful of representative files and pricing the rest against that. The inflow is the pile that matters, because it recurs for as long as you keep producing, and it is the one that turns a project cost into a line item.
Then there is the cost nobody quotes. Podcast teams with hundreds of episodes routinely pay for clean, human-corrected transcripts of the whole catalog, publish them, and then still end up running site: searches against their own website to locate a segment they half remember. The transcription was not wasted. It just bought them text, and text is not retrieval. If the reason you are transcribing is so people can find things, the budget has a second half that most plans leave out, and the work of getting from transcript to usable clip is where it goes.
This is also where the per-hour framing stops fitting. Vivu indexes each piece of material once when it is uploaded, and covers what appears on screen alongside what was said, so indexing is spent once per file and searches draw on a separate allowance. That does not replace a transcript you need for captions. It is a different purchase with a different unit.
When you should not be buying this at all
If you need five moments out of five recordings, scrubbing is cheaper than any quote you will get. If the transcript is the product, quality is the entire specification and the cheapest route is the wrong one by definition. And if you only ever look for spoken lines in recent material, the simplest form of spoken-word search may cover you without an archive-wide project.
Getting to a decision
Price one hour from a real file, not a clean sample, and read the output. That tells you the machine rate and the cleanup burden in one step. Multiply by the back catalog for the one-time number, by monthly output for the recurring one, and then ask what happens to the transcripts after they exist. If the answer is a folder of text files nobody opens, the per-hour rate was never the question you were trying to answer. If you are transcribing specifically to locate the good minutes inside long recordings, that job has its own shape and pricing it as transcription understates it.
FAQ
Is human-corrected transcription worth paying for if I only want to search the audio?
Usually not. Search tolerates errors that a reader will not: if a word comes through wrong in one sentence but right in the four other places the topic comes up, you still find the segment. Pay for human correction when someone reads the transcript as a document, when it feeds translation, or when it gets published. For retrieval alone, machine output plus a glossary of names and product terms gets you most of the way.
Are open-source speech-to-text models actually free?
There is no invoice, which is not the same thing. You pay in GPU time or CPU hours, in the setup, and in someone being responsible for the pipeline when it breaks. That trade tends to favor open models when the archive is large or the material is sensitive enough that you would rather not send it anywhere, and to favor a paid API when the volume is small and nobody on the team wants a new system to maintain.
How do I estimate the cost of transcribing a back catalog?
Take the total hours of audio, then transcribe a small sample first: your worst recording, your best, and something typical. The sample gives you a real error rate for your material rather than a marketing one, which is what determines whether you are buying machine output or machine output plus cleanup labor. Multiply the resulting per-hour cost by total hours, then add whatever the cleanup pass costs in staff time. That second term is the one that surprises people.
Does transcribing everything make my archive searchable?
It makes the words searchable, which covers dialogue, narration, and interviews. It does not cover anything that was shown and never spoken, and it does not by itself give you a way to search across hundreds of files at once. Plenty of teams end up with a complete, accurate set of transcripts and still no practical way to ask a question of the archive, because storing text and indexing it for retrieval are two separate steps.