# How to transcribe video to text > There are four practical routes, and the one you want depends on what a wrong word costs you. Canonical URL: https://vivu.ai/guide/how-to-transcribe-video-to-text There are four practical routes, and the one you want depends on what a wrong word costs you. The platform you already upload to will usually generate a transcript on its own. An open-source speech-to-text model such as Whisper runs on a machine you control and trades money for setup time. A transcription service takes the file and hands text back, with human correction available when the wording has to be right. And transcript-based editors produce text as a byproduct of editing, which is worth choosing only if you were going to cut in that tool anyway. ## Decide what the text is for before you pick A transcript you will skim yourself and a transcript that becomes published captions are different quality bars, and they justify very different effort. Machine output is generally fine for internal notes, for finding roughly where something was said, and for feeding into a summary. It is not fine, unedited, for anything with your name on it: proper nouns, product names, and anything said over music or crosstalk are where errors cluster. The honest test is to read two random minutes of machine output against the audio. If you would be embarrassed to publish those two minutes, you need a correction pass, and that pass is the expensive part of the whole job. ## The free transcript you probably already have Most video hosts and meeting platforms produce a transcript automatically once the file is up. This costs nothing and requires no decision, which is why it is the right starting point for almost everyone. The limits show up later. The text usually lives inside that platform's player rather than in a file you can move around, the timestamp granularity is whatever the platform chose, and a platform that re-encodes your upload is not a home for your master anyway. Export it if the platform lets you, and keep the exported file next to the video rather than inside the tool that made it. ## Running a model yourself Whisper and its variants are the standard open-source route, and they run locally, which means the audio never leaves your machine. You pay in setup: installing it, feeding it the right audio format, and waiting for long files. Extract the audio first rather than handing over the full video. Most people who go this way are transcribing a back catalogue in bulk, where a per-file charge would add up and a weekend of scripting will not. If you are transcribing one interview, this is the wrong route. ## Paying someone, and when correction is worth it Paid services bill by the length of the audio, and most offer a machine tier and a human-corrected tier. The second one exists because the last few percent of accuracy cannot be automated away, and it is the tier you want for anything legal, medical, or published as the record. What that ends up costing per hour is worth checking before you commit a large archive, and [the pricing shapes differ more than the tools do](https://vivu.ai/guide/how-much-does-speech-to-text-transcription-cost). Noisy source audio changes the math too, because it pushes machine output below the point where correcting is faster than retyping. ## What a transcript still does not give you Having text is not the same as being able to find things. Podcast archives with professionally corrected transcripts, published and public, still get searched with site: tricks in Google, because a pile of transcripts is a pile of documents and not an index. Ctrl-F only finds the words that were actually spoken, in the exact form they were spoken, so the moment where someone admits they were nervous is invisible unless they used your search word. Anything visual is invisible too, since nobody narrates what is on screen. A retrieval layer is the category that addresses this, and it works differently: the footage goes into a project, gets indexed once at ingest, and after that you ask for a moment in plain language instead of reading for it. [Vivu](https://vivu.ai/platform) is one of these, and a search for something like a speaker talking about nerves before a performance returns time ranges you can open, drawn from whichever videos contain them. It does not hand you a transcript and it does not cut anything, so it replaces the searching, not the transcription. ## When you do not need any of this If you have one video and you already know roughly where the part you want is, scrubbing is faster than any pipeline. If you need three pull quotes for a newsletter, play the file at 1.5x and type them. Transcription earns its place when the text is the deliverable, when captions are required, or when the volume is large enough that reading is no longer possible. Below that, the setup costs more than it saves. The way to sort yourself is to ask what you will do with the text five minutes after you get it. If you will publish it or edit from it, buy accuracy. If you will only ever search it, notice that you are solving a [search problem](https://vivu.ai/guide/search-video-by-spoken-words) with a transcription tool, and that the search problem has its own answers. And if you will edit from the text itself, you are looking for [a different category of tool](https://vivu.ai/guide/text-based-video-editor) than the one this question usually leads to. ## FAQ ### Can I transcribe a video without uploading it anywhere? Yes. Open-source speech-to-text models run locally on your own machine, so the audio never leaves it. You install the model, extract the audio track from the video, and run it. Expect to spend time on setup and to wait a while on long files, especially without a capable GPU. This is the route to take when the content is confidential or when you have a large batch and want to avoid per-file charges. ### Do automatic transcripts include timestamps? Usually yes, but at different granularity depending on the tool. Some give you a timestamp per segment or caption line, some give you one per word. Word-level timing is what you need if you plan to jump back into the video from the text or build captions that track speech closely. If you only ever plan to read the transcript, segment-level is fine. Check the output format before you process a whole archive, because re-running everything later to get finer timing is a real cost. ### Why is my transcript wrong on names and technical terms? Speech-to-text models predict likely words, and unusual proper nouns are by definition unlikely. Product names, people's names, and jargon specific to your field are the most common errors, and they are the ones readers notice. Some tools accept a custom vocabulary or hint list, which helps. Otherwise plan a correction pass focused on those terms rather than reading the whole thing: a find-and-replace on the handful of words you know it will get wrong usually fixes most of the damage. ### Should I transcribe the video file or the audio? Extract the audio. Transcription only uses the audio track, so handing over a large video file mostly costs upload time and disk space. A mono audio file at a modest sample rate is enough for most tools. The one thing worth doing first is checking whether the recording has music or two people talking over each other, because that, more than the file format, determines how much cleanup the output needs.