Turn an MP4 Into a Timestamped Transcript, SRT or VTT Subtitles

Get a transcript from your MP4 without uploading it

This MP4 to transcript tool listens to the audio in your MP4 and writes down what's said, on your own device. There's no upload and no sign-up. When it finishes, you can copy the transcript, read it on the page, or download it as a plain .txt file, an .srt subtitle file or a .vtt caption file.

The speech recognition runs on Moonshine, a fast open model from Useful Sensors, executed in your browser through WebAssembly. Moonshine ships as a separate model for each language, so you pick the language first and the tool loads the model trained for it. It covers English, Spanish, Japanese, Chinese and Korean.

How to convert MP4 to transcript

What people transcribe MP4s for

• Meeting and Zoom recordings. A call saved as MP4 becomes a document you can search, instead of scrubbing an hour of video to find one decision.
• Lectures and webinars. A recorded class becomes study notes you can search and review.
• Interviews and podcasts. A recorded conversation becomes the raw text a writer quotes from, or the base for show notes.
• Subtitles for your own videos. The SRT or VTT download gives you timed captions to load into an editor or a web player.
• Scripts and repurposing. A raw take turns back into text you can trim into a script, a blog post or social captions.
• Research sessions. User interviews and usability tests become transcripts you can code, quote and compare.

The tool gives you the words exactly as spoken, not a summary. That keeps it faithful to the recording, and you can summarize the text yourself or paste it into whatever tool you use next.

TXT, SRT or VTT: which download to pick

All three come from one transcription pass. Transcribe once, then pick a different format from the download menu if you need another. If you're unsure whether you need a transcript or subtitles, I compared them in video transcript vs. subtitles.

Languages and when to use the other tool

Because each language has its own model, the tool has a language picker instead of automatic detection. Pick the language actually spoken in your MP4. A model only decodes the language it was trained on, so the wrong choice gives you garbled or empty text instead of a fallback.

If your recording is in German, French, Italian or another European language, use my video to text converter. It runs a different multilingual model that covers ten languages without a picker, though it only downloads plain .txt. So the rule is simple: this page for subtitles and for English, Spanish, Japanese, Chinese or Korean; the other page for the ten European languages.

How the tool processes long recordings

The model reads audio in windows of about eight seconds rather than the whole file at once, because it's most accurate on short, bounded chunks. Each window is decoded on its own, without memory of the previous one. Now and then a word that falls exactly on the boundary between two windows comes out slightly wrong or split. Most sentences land cleanly, and it's the price of running a fast model on your own device instead of a big server model.

Accuracy depends on the audio more than the video. A close voice memo or a quiet meeting transcribes far better than a phone recording across a noisy room. People talking over each other is the hardest case for any transcriber, and music under dialogue lowers accuracy too. Use the result as a fast first draft for searching, skimming and notes, not as a certified word-for-word record.

There's no limit on file size or length. A ninety-minute recording is allowed the same as a ninety-second one; it just takes longer, because your processor does the work. Keep the tab open and in the foreground, and don't let a phone screen lock while it runs.

How the SRT and VTT timestamps are worked out

Moonshine doesn't output a timestamp for each word. It predicts text one token after another, without the frame-by-frame alignment some older speech models give. The only timing the tool knows for certain is where each eight-second window starts and ends.

To still give you usable subtitle cues instead of one long line per window, the tool splits each window's text into sentences at its punctuation. Then it gives every sentence a start and end time by spreading the window's eight seconds across its sentences in proportion to their length in characters. That assumes speech inside one window runs at a fairly steady pace. It usually does, but a window with a long pause or a burst of quick short words can drift a little. For captions people read along with a video, it syncs well. For frame-accurate professional subtitles, you'd still want a human timing pass.

The .txt download gets the same care for layout. Instead of one wall of text, the transcript breaks into a new paragraph wherever there's a pause of more than about a second, or once a paragraph reaches roughly five sentences. That's close to where a person listening would break it.

Frequently asked questions