# How much does it cost to transcribe and index 1,000 hours of video with cloud APIs?

> There is no single honest figure, because the bill is several lines metered in different units…

Canonical URL: https://vivu.ai/guide/how-much-does-it-cost-to-transcribe-and-index-1

There is no single honest figure, because the bill is several lines metered in different units, and the rates move often enough that a number printed on a page like this ages badly. What you can pin down is the quantities. A thousand hours is 60,000 minutes of audio to transcribe, on the order of eight to ten million words of transcript to embed if people talk the whole time, and somewhere between a few hundred thousand and a couple of million sampled frames if you want the picture searchable too. Multiply each by the current rate on your provider's pricing page, then add the lines no rate card shows: storage and transfer, search traffic, re-runs, and the person who keeps the pipeline working.

## Turn 1,000 hours into the units on the rate cards

Transcription is the easiest line to size. A thousand hours of video is a thousand hours of audio, or 60,000 minutes, unless you trim silence first. Check whether options you may want, such as speaker labels or a custom vocabulary for names, change the rate. How the per-hour side behaves is covered in [the per-hour arithmetic of transcription](https://vivu.ai/guide/how-much-does-speech-to-text-transcription-cost), and whether a machine pass will do depends partly on [how rough your audio is](https://vivu.ai/guide/what-s-the-best-way-to-transcribe-noisy-audio).

Embedding the transcript is the next line. Conversational speech runs at somewhere around 130 to 160 words a minute, so a thousand hours of steady talk comes to roughly eight to ten million words. Cut into passages of half a minute to a minute, that is 60,000 to 120,000 chunks, and each chunk becomes one vector.

Visual indexing is optional, and it is the line with the widest spread. At one sampled frame every ten seconds, a thousand hours is 360,000 frames. At one every two seconds, it is 1.8 million. That interval changes the frame count fivefold, so settle it before you compare any rates, and [the sampling trade-off](https://vivu.ai/guide/what-is-the-cheapest-way-to-generate-embeddings) is worth reading first.

Then storage. Every chunk and frame becomes a vector sitting in an index you keep paying to host, and the source files have to reach the provider, which can mean transfer charges and a second copy of the archive. Look up the real size of the thousand hours on disk, since bitrates differ too much between camera originals and web exports to guess.

## The lines no rate card shows

Having transcripts is not the same as having an index. A back catalog with clean transcripts posted online is searchable the way any website is, by words, and a hit names an episode page without saying where in the recording the answer is. The index is what maps a question to a position in a recording, and building it is separate work from the transcription.

Re-runs are next. Jobs fail and get retried, a vocabulary fix means re-transcribing the affected files, and switching embedding models means re-embedding everything, since vectors from different models can't share an index.

Search traffic is a running cost too. Every query gets embedded, and a reranking step or a generated summary adds more per question, so query volume belongs in the monthly estimate.

People are the last line. Someone writes the ingest job, handles retries, maps results back to timecodes and builds the screen where people check what came back. Put an honest hourly rate on that work and keep it in the same column as the invoices.

## Ways to assemble the same thousand hours

A speech API plus plain text search is the smallest build. You get exact phrase matching with timestamps, and nothing for what was only shown on screen.

A speech API plus embeddings, a vector index and your own interface covers paraphrase and, with frame sampling, the picture. The price is owning every line above.

Running open models on your own hardware removes the per-minute invoices and replaces them with compute and a person responsible for the machines.

A hosted search layer takes the build off your hands, and the upload becomes the step you plan around. [Vivu](https://vivu.ai/platform) is one of these: footage is uploaded to a project and indexed once in the cloud, so a new question doesn't reprocess the files. Searches count against its plan, which comes as a free tier and a paid tier, with the limits on Vivu's MCP page.

## When you don't need to index all 1,000 hours

Some of those thousand hours will probably never be searched. Index the shows and courses people actually request first, and add the rest if the searches earn it. If every question is about exact words, stop after transcription and text search. And if the archive gets searched a few times a month, a person with the transcripts open may cost less than any pipeline.

## How to read your estimate

Split the total into the one-time backfill and the monthly lines. If the backfill dominates, pick the route that gets the thousand hours through once without re-runs, and make the sampling and chunking decisions before anything is processed. If the monthly lines dominate (new footage, storage, query volume and the person keeping it running), the real choice is who operates the system, and that matters more than the per-minute rate on any pricing page.

## FAQ

### Is transcribing and indexing a video archive a one-time cost?

The backfill is, as long as nothing has to be redone. After that, new footage goes through the same steps as it arrives, the vector index costs something to host for as long as it exists, every search costs a little to run, and switching embedding models means re-embedding the whole archive. Budget the backfill as a project and everything else as a monthly line.

### Why do cost estimates for indexing video vary so much?

Because the inputs vary more than people expect. Whether visual indexing is included, how often frames are sampled, how many times the archive gets re-run, and whether anyone counted the engineering time can each move the total a long way. Two estimates for the same thousand hours are often pricing two different pipelines.

### Can I cut the cost by indexing only part of a video archive?

Yes, and it is the one saving that needs no negotiation. Start with the material people actually ask for, judged by past requests rather than guesses, and leave the rest as plain files until someone wants it. Transcribing everything but embedding only the most requested part is another middle path.

### What is the difference between transcribing video and indexing it?

Transcribing turns speech into text with timestamps. Indexing organizes that text, and optionally the picture, so a question can be matched to a position in a specific recording. A folder of transcripts can be searched for exact words and little else: a paraphrase won't match, and anything shown on screen isn't in the text.