SubtitleMedic

Auto-Generate Subtitles

Have a video with no subtitles? Drop it here and Whisper, a speech recognition AI, writes them for you: timed lines in SRT or VTT, ready to check and save. It runs on your own computer, so the file never leaves it.

Processed in your browserNo sign-up.mp4, .mov, .mkv, .webm, .mp3, .m4a, .wav

How to generate subtitles automatically

  1. Open the video or audio file

    Drop your file into the box above, or click to choose it. Videos (MP4, MOV, MKV, WebM) and sound files (MP3, M4A, WAV, OGG, FLAC) work, up to three hours long.

  2. Choose the model and the language

    Fast, Balanced or Accurate: bigger models get more words right but take longer and are a bigger download the first time. Leave the spoken language on Detect automatically, or pick it to help with short clips and accents.

  3. Generate, check and download

    Press Generate subtitles. The lines appear as each part of the file is heard. Read them through, choose SRT or VTT, and download the file, named after your video.

How automatic subtitles work

The page uses Whisper, a speech recognition model released by OpenAI and trained on hundreds of thousands of hours of speech in 99 languages. Whisper hears audio in windows of 30 seconds. For each window it writes the words it heard together with timestamps: where each stretch of speech starts and ends.

Before Whisper can listen, the sound is taken out of your file and turned into the form the model expects: a single channel, 16,000 samples per second. A film's sound is read in blocks of about ten minutes, each cut at a pause so no word is split in two, so even a three-hour recording never has to sit in memory at once.

Whisper's stretches of speech don't make good subtitles on their own: some last 20 seconds, some hold a whole paragraph. So each stretch is cut into subtitle-sized lines, at the end of a sentence where possible, then at a comma, then between words. Each line gets its share of the stretch's time, is wrapped onto at most two rows of 42 characters, and stays on screen for at most seven seconds, the limits most streaming services use. Lines that hold no words, and Whisper's habit of repeating a phrase over silence, are left out.

Everything runs on your computer

Online subtitle generators usually upload your video to a server, run the recognition there and send the text back. This page doesn't. The model is downloaded to your browser and runs there, through WebGPU on your graphics card when the browser offers it, and through WebAssembly on your processor otherwise. That matters for interviews, meetings, lessons and family videos you would rather not hand to a stranger's server, and it means there is no upload to wait for, no account and no minutes limit.

The model files come from Hugging Face and the runtime from jsDelivr, both at fixed versions; the runtime is checked against a known fingerprint before it runs. Nothing of yours is sent to either.

Choosing the right model

  • Fast (Whisper tiny) is the smallest download and the quickest, even without WebGPU. It is fine for clear speech in English and for a rough transcript, and it struggles with accents, noise and less common languages.
  • Balanced (Whisper base) is the default when your browser has WebGPU. It is noticeably better at names, punctuation and other languages, at about twice the time.
  • Accurate (Whisper small) is offered only with WebGPU, because on a processor alone it would take longer than the video. Choose it for Vietnamese, Indonesian and other languages Whisper heard less of, or for a recording with background noise.

The badge Downloaded marks the models already in your browser: using them again costs no download at all.

Getting the best result

Clean sound helps most. Whisper does best with one person speaking clearly near the microphone. Music under the voice, a noisy room and people talking at the same time all cost accuracy. If you can, use the original recording rather than a video with a soundtrack added.

Set the language for short clips. Automatic detection listens to the first 30 seconds. A clip that starts with music, or with a few words in another language, can be detected wrongly, and then the whole file comes out in the wrong language.

Expect to edit. Even a good result needs a read-through: names spelled the way they sound, a missed word in a fast exchange, a line that belongs to the next speaker. Correcting a draft is still much quicker than typing everything yourself.

What to do with the subtitles next

  • Play them with your video in VLC, Plex or a TV by giving the SRT the same name as the video, as explained in how to add subtitles to a video.
  • Put them inside the file as a track viewers can turn on and off with Add Subtitles to MP4.
  • Draw them into the picture for Instagram, TikTok or a presentation with Burn Subtitles into Video.
  • Upload the SRT or VTT to YouTube or Vimeo as your own caption track, which replaces their automatic captions.

Automatic captions are not the same as subtitles for the deaf and hard of hearing: they write down speech but not sounds such as music or a door slamming. The difference is explained in what SDH subtitles are.

Frequently asked questions

How do I automatically generate subtitles for a video?

Drop the video into the box above, pick how accurate the model should be and the language that is spoken, and press Generate subtitles. Whisper listens to the sound and writes timed lines, which you can read through on the page and download as an SRT or VTT file.

Is my video uploaded to a server?

No. The sound is read and recognized inside your browser, on your own computer. The only thing downloaded is the speech model itself, once, from Hugging Face; your file, its sound and the words found in it never leave your device.

How accurate are the subtitles?

With clear speech and little background noise, Whisper gets most words right, and each bigger model makes fewer mistakes than the one before. Names, technical words, quiet voices and people talking over each other or over music are where it slips, so always read the lines through before you publish them.

Which languages can it recognize?

Whisper was trained on 99 languages, including English, Spanish, Vietnamese, Indonesian, Chinese, Japanese, Hindi and Arabic. It does best on the languages it heard most, so English, Spanish and other widely spoken languages come out cleanest. Pick the language yourself for short clips or strong accents; automatic detection decides from the first 30 seconds.

Why is it slow on my computer?

The model runs fastest on the graphics card through WebGPU, which Chrome and Edge offer on most computers. Without it, the page uses the processor, which can take several minutes for a long video. Choose Fast for a quicker result, close other tabs, and keep the tab open until it finishes.

How big is the download?

Fast is 45 to 125 MB, Balanced 80 to 210 MB and Accurate, offered only with WebGPU, 410 to 590 MB, depending on what your graphics card supports; the page shows the exact size beside each model. It is downloaded the first time you use a model and kept in your browser, so the next video starts at once. A button under Generate removes the downloaded models when you need the space back.

Can it translate the subtitles into another language?

No. This page writes down what is said, in the language it is said. To get subtitles in another language, generate them here first and then translate the SRT file, for example in a spreadsheet with SRT to Excel.

Does it work on a phone?

It can, for short clips. Phones have less memory and slower processors, so a long video may stop halfway. On a phone, choose Fast and keep the clip short, or use a computer for anything longer than a few minutes.