That question is two jobs, and the answer is different for each. Transcription has good options you can run without a vendor, mostly open-source speech-to-text models and the transcripts that fall out of software you already use. Search over those transcripts is where the free path runs out. What you end up holding is a folder of text files, and the people who get furthest with this discover that having 300 clean transcripts and being able to find anything in them are unrelated achievements.
The transcription half
Open-source speech recognition models are the category to look at if you are willing to run something. Whisper is the best-known family. You run it on your own machine or a rented one, which means the constraint is your hardware and your patience rather than a per-minute meter, and it means the output stays where you put it. This is the route for a back catalogue, because a batch that takes two nights costs you two nights of an idle laptop.
Transcripts you already have are the option people overlook. Video platforms, meeting software, and editing suites generate them as a byproduct, and a lot of archives are sitting on caption sidecars from a captioning job done years ago. Before commissioning anything, go look at what your files already ship with.
Human-corrected transcription is a different product from either of these. It exists because machine output has a floor on proper nouns, crosstalk, and domain vocabulary that no amount of reprocessing fixes. Groups that publish transcripts as a deliverable, rather than using them internally, usually end up paying for the correction pass whether they planned to or not.
The thing that separates these routes is not really quality; it is who fixes the errors, and how much your material has of the kind of speech machines get wrong.
The search half, which is where it breaks
A transcript makes a video greppable, and greppable is a much lower bar than searchable. Two things go wrong almost immediately.
The first is that people search for what a segment was about, not for the words in it. Nobody remembers the exact phrasing from an episode recorded three years ago; they remember the argument. Text matching returns nothing for a paraphrase, which is why searching an archive you transcribed yourself so often feels like the transcripts are not working when the transcripts are fine.
The second is that a hit in a text file is not a place in a video. You get a document with the phrase in it, then you open a two-hour file and hunt. Formats matter here: SRT and VTT carry timecodes, plain text does not, and if you transcribed to plain text you have thrown away the one thing that would have taken you back to the moment.
The free tooling for this half is what you can assemble: site-restricted searches against a published site, grep across a directory, a search plugin on whatever the transcripts are hosted in. All of it works up to a point and none of it survives a few hundred hours.
There is also the category of question a transcript cannot answer at all, which is anything nobody said out loud. This is where a retrieval layer differs from a transcript index: Vivu searches the footage itself and returns the matching time range, so the shot where the product is actually on the table comes back even though no one narrated it. Words remain the easier surface, and the mechanics of searching by them are covered in finding video by what was said in it.
What free actually costs here
Machine time, which is usually fine. Cleanup, which is not: proper nouns and speaker labels are where the hours go, and they are exactly the terms people search for. Then the maintenance nobody assigns, because a transcript archive is only useful if every new episode gets added, and the day that stops happening is the day people go back to asking a colleague.
When you do not need a tool for this
If you have a handful of videos, watch them. If you need one quote from one file, scrub for it. If the archive is dormant and nobody has asked a question about it in a year, transcribing it is a project with no reader. The threshold where this becomes worth building is somewhere around the point where more than one person needs to find things in material they did not personally make, which is also the point where building an actual content library becomes the real task rather than a side effect.
Which problem you actually have
If you have no transcripts, start with the open-source route on a sample of ten files, look at the errors, and decide from there whether correction is a cost you need to absorb. If you already have transcripts and still cannot find anything, more transcription will not help you, and neither will a better free transcriber. The gap is retrieval, and it is a different purchase, a different build, and a different question than the one most people start with. The cost side of transcription specifically, if that is still the open decision, is laid out in what transcription costs per hour.
FAQ
Can I search a transcript for something that was shown but not said?
No. A transcript only contains speech, so anything visual that nobody described is invisible to it: a location, a product on screen, a reaction, a slide nobody read out. This is the single most common reason a transcript archive disappoints the person who built it. Finding those moments requires an index built on the footage itself rather than on the audio track.
Do free transcription tools give me timestamps?
It depends on the output format you choose, not on the tool being free. SRT and VTT carry timing information, JSON output from most speech recognition models carries it at word or segment level, and plain text throws it away. Pick a timed format even if you only want to read the text today, because regenerating the transcript later to recover timestamps means running the whole archive again.
How accurate does a transcript need to be for search to work?
Accurate on names, products, and jargon; forgiving everywhere else. Ordinary words being slightly wrong rarely affects whether you find a segment, because you are usually searching a distinctive term. A misspelled company name or a mangled speaker label makes an episode effectively unfindable. If you are correcting transcripts by hand, correcting only the proper nouns gets you most of the search value for a fraction of the work.
Is it worth transcribing an archive nobody searches yet?
Usually not as a bulk project. Transcribing on demand as questions come up costs almost nothing and tells you what people actually ask, which is the information you need before deciding how to index. If the questions never come, you saved the whole project. If they arrive weekly, you now have a real specification instead of a guess.