Transcribe to text
Speech to subtitles or a plain transcript, without sending the recording anywhere.
Your file
Drop a file here, or click to choose
It stays on this device. Nothing is sent anywhere.
Settings
Check the transcript
What you get
- SRT is what YouTube, Vimeo, Premiere and almost every player accepts. Upload it alongside your video and viewers get switchable captions.
- VTT is the same idea for web players and the HTML
<track>element. - Plain text is the transcript with no timings, for notes, show notes, blog posts and search.
Timings come from word-level timestamps, so lines break where speech actually pauses rather than on a fixed clock.
Editing before you publish
No speech recognizer is perfect on names, jargon or crosstalk. SRT is plain text, so open the result in any editor, fix what is wrong, and either upload it as-is or burn it into the picture with the subtitle burner. That round trip is the reliable way to get captions that are actually correct.
Podcasts, interviews, lectures
Length is not limited by anything except your machine's memory and your patience: transcription runs at roughly three to six times real time, so an hour-long interview takes ten to twenty minutes. Because none of it is uploaded, this is the option that works for recordings you are not allowed to hand to a third party, like client calls, medical or legal material, and unpublished research.
The confidentiality case, specifically
This is the reason the tool exists in this form. Transcription services are convenient and they work by receiving your recording, which for a great deal of material is exactly the thing you cannot do. A solicitor's client interview, a doctor's dictation, an HR investigation, a therapy session, a journalist's conversation with a source, an unpublished research interview under an ethics approval: in every one of those cases sending the audio to a third party is either a breach of duty or a breach of contract, regardless of what that party's retention policy claims.
Here the model runs in the tab. The weights are downloaded once from Hugging Face and cached, which tells that host your IP address and which model you asked for and carries none of your audio. After that first download the tool works with the network disconnected, which is the simplest way to satisfy yourself that the recording is not going anywhere.
Recording quality decides accuracy more than any setting
Nothing improves a transcript as much as better audio, and no option here compensates for a bad recording. What helps, in order: a microphone close to the speaker rather than across a room; one person talking at a time, since crosstalk defeats it entirely; a quiet room, because background music and traffic are worse than they sound; and the original file rather than a compressed copy, since a heavily compressed voice recording has already lost detail the model needed.
Where it reliably struggles is proper nouns, brand names, technical vocabulary, acronyms and heavy accents. It will also, occasionally, produce a completely fluent sentence that nobody said, usually during silence or noise. That is worth knowing because it does not look like an error; it looks like text.
Read it before you use it
The output is plain SRT or text, and correcting it in any editor is far faster than typing from scratch. On anything that matters, read the whole thing against the audio rather than spot-checking. Names and numbers are where errors concentrate and where they do the most damage.
Once corrected, the transcript can be uploaded as a caption file wherever that is supported, or burned permanently into the picture with burn subtitles. Auto captions collapses both steps into one when the text needs no editing, and the editor lays captions on a timeline as individual lines you can restyle and retime.
Common questions
Is there a length limit?
No hard limit. The audio is held in memory, so very long recordings can strain a phone. On a desktop, an hour is comfortable.
How accurate is it?
On clear speech, the accurate model is close to a careful human transcript apart from proper nouns. Noisy or overlapping speech is harder, as it is for every recognizer. Always read it through before publishing.
Does it identify who is speaking?
No. Speaker separation is a different model than the one running here, so a two-person interview comes out as one continuous transcript.
Can it translate?
It can translate speech in another language into English subtitles: pick the spoken language and tick translate. It does not translate English into other languages.