How to transcribe a video to text, and what to check afterwards

There are three ways to transcribe a video to text: listen and type it out by hand, use the automatic captions of the platform where the video already lives, or run it through an automatic transcription tool, which returns the full text and timed subtitles within minutes. For people who record for work — webinars, lessons, client calls — the third is the one that holds up, followed by a targeted read-through of the spots where the machine goes wrong: proper names, technical terms, figures.

This guide compares the methods, says where each one goes wrong and describes the workflow worth setting up once and for all, with a real example taken from our demo.

From video to text: the methods compared

By hand

You listen, pause, type, rewind. As an order of magnitude — a common estimate, not a measurement of ours — one hour of recording takes three to five hours of listening and typing, and more when many people talk or the audio is poor. For anyone running a weekly webinar, that's half a working day or more, every week.

It remains the right choice in a few precise cases: when the text must be faithful word for word and someone has to answer for it, like minutes of a meeting, or when the audio is so difficult that correcting a draft would cost more than rewriting it. The full reckoning, with the cases where a person does better than the machine, is in the comparison with manual transcription.

The platform's automatic captions

If the video is already on YouTube, the platform often generates automatic captions by itself, and the channel owner can review and edit them in YouTube Studio. Some video-conferencing platforms also produce a transcript of the meeting, depending on the plan and the settings.

They are handy when they're there, with two limits. They aren't always there: they depend on the language, the audio quality and how the account is set up. And they live where the video lives: a lesson reserved for enrolled students or a call with a client doesn't go on a public platform.

Your computer's dictation

Some people play the recording into the operating system's dictation. It can be done, but dictation is designed for a person speaking into the microphone, not for audio played through the speakers; it has to run in real time, so an hour of video means an hour in front of the computer; and it returns text without timings, so no subtitles.

An automatic transcription tool

You upload the file, or paste a link, and a speech recognition engine produces the text of everything that was said, with the timing of each sentence. Subtitles come from the timings; everything else comes from the text. The machine's work takes minutes; what's left for you is the read-through.

And the read-through is needed, because automatic transcription goes wrong in predictable places:

• Proper names and brands. An uncommon surname turns into a word the model knows, and the sentence still flows: the error reads like correct text.
• Technical terms and acronyms. Your industry's jargon is the first thing to slip, especially when it's said quickly because the speaker takes it for granted.
• Overlapping voices. Audience questions, interruptions, two people talking at once: it's the hardest case, harder than background noise.
• Poor audio. A distant microphone, room echo, the voice of someone joining remotely on a shaky connection.
• Figures. A number said aloud may come out in words or in digits, and a wrong one goes unnoticed when you read.

How much these errors weigh, and how to measure them on your own files in two minutes, is the subject of Automatic subtitles: what to check before you trust them.

The workflow that holds up

If you transcribe once in a while, any method will do. If you record every week, a three-step workflow pays off, the same every time.

1. Clean audio, before you record

It's the step that costs least and pays most. For the text, the microphone matters, not the camera: one close to the mouth is worth more than any framing.

• Webinars: the moderator repeats audience questions aloud before answering. A question asked into a distant microphone gets lost; repeated, it makes it into the text.
• Classroom lessons: the microphone is clipped to whoever is teaching, not left on the desk, and students' questions are repeated as in webinars.
• Calls: headsets with a microphone for everyone, and one person speaking at a time when it comes to decisions.

The consent to record a call is yours to ask for, before you press the button.

2. Automatic transcription of the original file

Upload the original recording, not a recompressed copy passed around through a chat app: every extra round of compression takes something away from the voice. If the platform lets you download the audio only, that works too, because for the text that's all that matters.

3. A targeted read-through, not a full one

Rereading an hour of text with the same attention from start to finish is the slowest way to find the errors, and the easiest way to miss them. It pays to look for them where they are:

1. Before you start, write the list of names: speakers, guests, companies, products, industry acronyms. They are where the machine gets it wrong most.
2. Search for them in the text one by one, with the editor's search. A mangled variant jumps out next to the correct one.
3. Check the figures: prices, dates, percentages. They're the words that do the most damage when wrong.
4. Listen again to the multi-voice passages: the Q&A, the interruptions. That's where an error is most likely and hardest to guess from the text.

The rest gets read once, at normal speed. And if you turn the text into an article or a newsletter, check the same names there too: an error in the transcript can carry over into the texts built from it.

A real example: what we corrected in our demo

The echoflow public demo is a video actually processed by the product: just over five minutes, a single synthetic voice reading a script, no noise. It's the easiest case there is. Before publishing it we checked every piece against the script, and in the transcript we found one error to fix: where the voice says "eight hours", the transcript read "eight to eight minutes. hours".

That's the broken kind of error, the one anyone spots on a first read. The Italian edition of the demo had the other kind, and it's the more insidious one: where the voice says "sensi di colpa" (guilt), the transcript read "sensi di colta" — a real Italian word, which no Italian spell-checker flags and which reads without a stumble. You only find it by rereading with the script, or the recording, in front of you. Those are the two families to look for.

The same transcript also shows a flaw that isn't an error: figures aren't consistent. In the same text there's "5 practical habits" in digits, but "Five minutes" and "Ten minutes" in words. In a subtitle it changes nothing; in a text you're going to publish, making them consistent is part of the read-through. What you see in the demo today is the reviewed version, with these flaws of form left where they are.

Here is the same sentence from the demo in the two main formats. In the transcript it's continuous text:

You are not short on time. You are short on attention.

In SRT subtitles it becomes a numbered block, with a start time and an end time:

3
00:00:05,400 --> 00:00:08,946
You are not short on time.
You are short on attention.

Which format you need, and what for

• Text (.txt): to read, search for a sentence, quote a passage or turn it into an article. It's continuous, without timings and without the names of who is speaking.
• SRT: the subtitle format accepted by almost every player, video platform and editing program. Each block has a number, a start time and an end time.
• VTT (WebVTT): the same principle, in the native format of web video players. It's what you need when the video sits on a page you manage.
• Video with burned-in subtitles: for social media, where many people watch without sound. Burned-in means they are part of the image and can't be removed.

SRT and VTT are text files: a wrong name can be fixed in any editor, without touching the timings.

What to do with it once you have it

The transcript is rarely the end point. It's the raw material for something else:

• Subtitles, for people watching without sound and for people who can't hear: automatic subtitles.
• An article, which is what the video can't be: linkable, indexable, quotable — from video to blog article.
• Chapters with timings, for whoever rewatches a long recording looking for the part they need: chapters and show notes.
• Course materials from a lesson: summaries, chapters and, with Pro, the lesson handout — for trainers.
• The recap for those who missed it and a week of content from a single webinar: for teams that run webinars.

With echoflow, in practice

Upload an audio or video file, or paste a YouTube, Google Drive or direct file link; for a podcast the RSS feed is enough. Within minutes you have the transcript in the language actually spoken, SRT and VTT subtitles and, from the same upload, a summary, chapters, an article, a newsletter and posts: all drafts, to be read through as above. Every text can be corrected on the video's page, and what you download is the corrected version. Formats and details are on the automatic video transcription page.

What it doesn't do, said up front:

• It doesn't recognise who is speaking: in an interview, you add the names in front of the lines yourself.
• It doesn't join meetings: there's no bot dialling into the call, you upload the recording afterwards.
• It has a length cap per video: 60 minutes in the trial, 180 on Pro. A longer webinar is uploaded in two parts.

The file you upload is deleted within 24 hours of processing, unless the workspace has set its own retention; generated media files, such as the captioned video and the clips, stay downloadable for 7 days; text doesn't expire while the account stays active. Files and content are stored in the EU; transcription and AI generation also go through non-EU providers, listed by name on our sub-processors page, under the safeguards of Article 44 and following of the GDPR. Worth knowing before you upload a call with a client.

The trial lasts 14 days and gives you 60 minutes, up to 10 videos, no card; plans and minutes are on the pricing page. The advice is the one above: upload a typical recording, with its questions and its difficult names, not the cleanest one you have.