Transcribe video to text

What a transcript is good for

Turning speech into text pays off whenever you need to search, quote, skim or reuse what was said rather than listen to it again. Typical uses:

• Interviews and research calls, where you need exact quotes with a rough idea of where they were said.
• Lectures and webinars, so you can search the text for a term instead of scrubbing through an hour of video.
• Video subtitles for social media, where most viewers watch with the sound off.
• Meeting recordings and podcast episodes that need notes or an accessible text version.

A plain transcript is the best choice when you will read or edit the words. SRT or VTT is the choice when the text has to appear on screen in sync with the video.

How recognition works here

Speech recognition runs on our servers, not in your browser. The engine is Whisper, a speech model published by OpenAI in 2022 and trained on roughly 680,000 hours of multilingual audio, run through the faster-whisper implementation.

What happens to your file:

1. The file is uploaded and its duration is checked.
2. A voice activity detector skips stretches of silence and pure background noise. This speeds things up and stops the model from inventing words in quiet passages.
3. The model works through the audio in short windows and produces text segments with start and end times. It adds punctuation and capitalisation on its own.
4. The segments are written out in your chosen format: one segment per line for TXT, numbered cues with 00:01:02,500 style timings for SRT, and the same with 00:01:02.500 timings under a WEBVTT header for VTT.
5. The uploaded media is deleted as soon as the result is written. The job folder with the result is removed automatically within about an hour.

Processing speed is close to real time: a 10-minute recording needs roughly 10 minutes.

Choosing language and format

The spoken-language menu offers automatic detection plus English, Russian, Ukrainian, Polish, German, French, Spanish, Portuguese, Italian, Turkish, Indonesian, Vietnamese, Japanese and Chinese.

Detection listens to the first part of the recording and picks the most likely language. It is reliable for a clear single speaker. Set the language yourself when:

• The recording starts with music, a jingle or silence.
• Speakers have a strong accent; detection can mistake accented English for the speaker's native language and transcribe in the wrong one.
• Russian and Ukrainian, or Spanish and Portuguese, are mixed or easily confused.

For format, SRT is the safe default for subtitles: YouTube, video editors and desktop players all read it. Pick VTT if the subtitles will go into an HTML5 <track> element on a website. TXT is the right choice for notes, quotes and articles; the text also appears on the page, ready to copy.

If you need the speech in English rather than in the original language, use translate video, which runs the same engine in translation mode.

Getting a cleaner transcript

Audio quality decides most of the outcome.

• One microphone per speaker beats one microphone for the room. A laptop mic three metres away picks up echo, and echo is the most common cause of dropped or merged words.
• Background music hurts. Speech over a music bed is transcribed noticeably worse than the same speech alone. Use the raw recording when you have it.
• Crosstalk is not separated. When two people talk at once, one voice usually wins and the other is lost. The output also does not label who is speaking; add speaker names while you proofread.
• Check names, numbers and jargon. Proper names, product names, acronyms and figures are where errors cluster. Search the transcript for them before publishing.
• Expect long subtitle cues from fast speakers. Cue length follows natural phrases; a subtitle editor can split them.

Limits and how to work within them

Free use allows one server run a day per device without an account, for recordings up to 10 minutes and files up to 500 MB. Premium removes the daily limit and accepts recordings up to 90 minutes per file.

Ways to stay inside the limits:

• Video is mostly picture. A 10-minute 1080p screen recording can weigh 300 MB or more, while its audio track as an MP3 is about 10 to 15 MB. Extract the audio first and upload that instead: the upload finishes faster and the transcript is identical.
• For longer recordings, cut them into parts of 10 or 90 minutes and process each one. Split at a pause rather than mid-sentence so no words are lost at the boundary.
• Accepted inputs are MP4, MOV, MKV and WebM video, and MP3, WAV, M4A, OGG and FLAC audio.