Is my recording uploaded?

No. The audio is decoded in your browser and transcribed by a Whisper model running on your device. Only the model file is downloaded, once, and then cached.

How accurate are the subtitles?

Clear, single-speaker speech is usually accurate. Tiny is the fastest option and Base is noticeably better. Background noise, music, overlapping speakers, and strong accents increase errors, so review the cues before publishing.

Which file types can I use?

MP3, M4A, WAV, OGG, Opus, and FLAC audio, plus the audio track of MP4, MOV, WebM, and MKV files your browser can decode. Video is never uploaded or re-encoded; only its soundtrack is read.

Does it detect the language automatically?

Yes. The model identifies the spoken language from the audio, so English, Spanish, French, German, Hindi, Japanese, and many others work without extra settings. Short clips and heavy accents can be misdetected.

How long can a recording be?

Up to 30 minutes per run. Longer recordings need more memory and take much longer than real time, so split them with the Video Trimmer or an audio editor first.

Can I edit the subtitle timing?

Yes. Download the SRT, or select Edit in Subtitle Editor to load the cues straight into the editor and adjust wording, timing, or line breaks before you burn them into a video.

Auto Subtitles — Transcribe Video & Audio

What Auto Subtitles helps you do

Turn a recording into a subtitle file without sending it anywhere. A Whisper speech-recognition model runs in your browser, detects the spoken language, transcribes the audio, and groups words into timed cues you can download as SRT or refine in the Subtitle Editor. It is aimed at short clips and screen recordings, not hour-long archives.

Audio is decoded and transcribed in your browser by a Whisper model loaded with transformers.js. Your recording is never uploaded; only the model file downloads once and is cached by the browser.

How it works

The browser decodes the audio and feeds it to a Whisper model loaded with transformers.js. The model predicts the spoken language and a stream of words with timestamps, which the page groups into cues. The model file is downloaded once and cached; the audio itself is never uploaded.

How to use Auto Subtitles

Drop an audio or video file, or browse to select it from your device.
Choose the Tiny model for speed or Base for accuracy, then select Generate subtitles.
Review the transcript against the audio for names, numbers and technical terms.
Download the SRT, or select Edit in Subtitle Editor to adjust the cues.

When to use it

Add captions to a screen recording, tutorial, or interview.
Turn a voice memo or meeting recording into searchable text.
Draft subtitles for a social clip, then correct names and timings before publishing.
Create a transcript for a short product demo.
Caption a lecture clip for students who need to read along.
Draft a translation source from a spoken clip.

When not to use it

Do not use it for recordings longer than 30 minutes without splitting them first.
Do not publish the raw transcript without reviewing names, numbers and technical terms.
Do not use it for confidential audio on a shared device without consent.

Example

An 11-second speech clip is dropped in with the Tiny model. The tool detects English and returns the transcript plus 22 word timestamps grouped into readable cues, ready to export as SRT. A longer interview is split into ten-minute sections first so each run stays within the memory limit.

What you can do with Auto Subtitles

Transcribe audio or the soundtrack of a video file.
Detect the spoken language automatically.
Choose the Tiny model for speed or Base for accuracy.
Group word timestamps into readable cues.
Download the result as an SRT file.
Send the cues straight to the Subtitle Editor.

Example: caption a short tutorial

Trim the recording to the section that needs captions, then drop the file into the tool. A short clip finishes faster and is easier to check line by line.

Start with the Base model when accuracy matters, and review names, numbers and technical terms against the audio. The transcript is a draft, not a final caption file.

Download the SRT or open it in the Subtitle Editor to fix timing and line breaks, then burn the captions into the video or upload the track to the platform.

What to know before you start

Tiny is fast but less accurate than Base; review names, numbers, and technical terms.
Recordings are capped at 30 minutes because the whole clip is processed in memory.
Processing runs slower than real time on CPU-only devices; WebGPU speeds it up where available.
Music, crosstalk, and heavy background noise reduce accuracy.
Cue timing is grouped from word timestamps, so line breaks still need a pass in the Subtitle Editor.

Common mistakes

Choosing Tiny for a noisy interview and accepting the errors.
Assuming the detected language is always right for a short or heavily accented clip.
Publishing SRT cues without checking line length and timing.
Running a 40-minute file and losing the tab to memory pressure.
Treating the transcript as a verbatim legal record.

Troubleshooting

If the language is wrong, try a longer or cleaner section; the detector needs enough speech to decide.

If the transcript is missing words, the audio may be quiet or clipped. Normalize the source or choose the Base model.

If processing is very slow, the device is running the model on CPU. Try the Tiny model or a shorter clip; WebGPU helps where available.

If the run fails on a long file, split it with the trimmer and process each part separately.

If the SRT timing drifts, review the cues in the Subtitle Editor and shift the file as a whole if needed.

If the first run downloads a model and then stops, the browser may have blocked the model request or run out of memory. Reload and try a shorter clip.

Auto Subtitles