Audio to Subtitles Converter

Generate SRT or VTT subtitles from speech, without uploading the audio.

Runs in your browser Free · no accountFile upload supported
Downloaded once, then cached by your browser for future visits.
42 is the broadcast convention. Longer lines are harder to read at a glance.
Multilingual model only. Transcribes non-English speech as English text.

File

Result

Your converted file will appear here.

This conversion happens locally in your browser. Your files are never uploaded.

About the Audio to Subtitles converter

Speech recognition normally means uploading your recording to somebody else's server and waiting in a queue. This converter runs the model on your own machine instead: the audio never leaves the browser tab.

Drop in a recording and get back timed subtitle cues in SRT or WebVTT, or a plain transcript if you only want the words.

What to expect from the accuracy

Clear single-speaker audio — a lecture, a podcast, a voice memo — transcribes well, and the base model is noticeably better than tiny at proper nouns and technical terms.

Accuracy falls with background noise, strong accents, crosstalk and music. The model does not label speakers, so an interview comes out as continuous text rather than a dialogue.

Treat the output as a strong first draft. For anything published or legally significant, read it against the audio before you rely on it.

Choosing a model

Base is the recommended starting point at around 82 MB — accurate enough for real work and quick to download. Tiny is half the size and noticeably rougher, which is a fair trade on a slow connection or for a rough draft.

The multilingual model is larger and handles 90-plus languages, including translating non-English speech into English subtitles. Pick an English-only model when your audio is English: they are smaller and more accurate at the same size.

Subtitle formatting

Cues are wrapped to two lines at 42 characters, which is the broadcast convention and roughly what a viewer can read in the time a line is on screen. Overlapping segments are trimmed so each cue ends before the next begins, since many players reject files that overlap.

Very short segments are padded to a minimum duration so a one-word cue does not flash past unreadably.

How it works, and where the model comes from

Your browser decodes the audio, resamples it to the 16 kHz mono that Whisper expects, and hands it to the model running in a background worker. Nothing is uploaded, so no server ever holds your recording.

The model itself is a few dozen megabytes and is downloaded the first time you use the tool, then kept in your browser's cache. Later visits start immediately and work with no network connection at all.

Where WebGPU is available — Chrome, Edge and recent Safari — recognition runs on the graphics card and is typically faster than real time. Elsewhere it falls back to the CPU, which works but is several times slower.

Frequently asked questions

Is my recording uploaded anywhere?

No. The model runs inside your browser tab, so the audio never leaves your device. The only network request is the one-time model download, and self-hosted model files mean even that does not reveal what you are transcribing.

Do I have to download the model every time?

No. It is stored in your browser's cache after the first use, so later visits start straight away and work offline. Clearing your browser's site data removes it, and it downloads again next time.

How long a file can I transcribe?

Long files work — the audio is processed in 30-second windows with overlap — but it takes real time on your own hardware. An hour of audio is comfortable with WebGPU and slow without it. Leave the tab open while it runs.

Which languages are supported?

The English-only models handle English. The multilingual model covers over 90 languages and can also translate them into English subtitles.

Can it tell speakers apart?

No. Speaker separation is a different task from recognition, and Whisper does not do it. The transcript is continuous text.