BRev Transcribe audio & video — free & private
Transcribe audio and video to text with OpenAI Whisper in your browser — pick from 36 languages or auto-detect any of Whisper’s 99, SRT/VTT subtitles, never uploaded.
This transcription tool runs OpenAI's Whisper speech-recognition model inside your browser tab, so the recording is decoded and transcribed on your own device and never sent to a server. You can choose an audio or video file, or record straight from your microphone. There are three model sizes: Speed is Whisper tiny and English only, Balanced is Whisper base and the default, and Accurate is Whisper small, slower and more accurate. On the CPU they are downloads of about 40, 80 and 250 MB; when your browser offers WebGPU the tool loads higher-precision weights, which are larger. The model files come from Hugging Face the first time and stay in your browser's cache. Balanced and Accurate transcribe the 36 languages in the Language list, Arabic, Urdu and Hindi among them, and on Auto-detect they recognise any of the 99 languages Whisper was trained on from the first 30 seconds; they can also translate that speech into English. With Segments on, you get timed lines to save as SRT or VTT subtitles or as JSON, and the plain transcript is editable and saves as a .txt file.
What you can do
- Three Whisper models: Speed (tiny, English only), Balanced (base, multilingual, selected by default) and Accurate (small, multilingual)
- Choose an audio or video file, drop one on the page, or record from your microphone with “Record”
- Pick the spoken language from a list of 36 that includes Arabic, Urdu, Hindi, Persian, Tamil, Bengali and Swahili, or leave it on Auto-detect; use “Translate → EN” for an English transcript
- Timed segments saved as .srt or .vtt subtitles or as .json, with a line-length setting (42 characters by default, 0 for none) that moves longer captions onto a second line
- Runs on the GPU through WebGPU when the browser provides it, and otherwise on the CPU with an 8-bit model; the result shows which one was used
- Editable transcript with Copy, Copy with times and a .txt download named after your file
- No account and no per-minute charge; the model and its runtime are downloaded, your audio is not sent anywhere
How to use audio & video to text
- Click “Choose audio / video”, or switch to “Record”Under “Audio source”, “Choose file” is selected: pick an audio or video file, or drop one on the page. To use your microphone, switch to “Record”, press “Start recording” and then “Stop recording”.
- Under “Model”, pick “Speed”, “Balanced” or “Accurate”Balanced is selected by default. Speed is English only; for any other language keep Balanced or Accurate, choose the spoken language from the “Language” list, and set “Task” to “Translate → EN” if you want English text.
- Choose “Segments” or “Plain” under “Timestamps”Segments is the default and is required for SRT and VTT subtitles; Plain gives text only. “Advanced” holds the subtitle line length, 42 characters unless you change it.
- Click “Transcribe”The first run shows the one-time model download, then “Transcribing on GPU…” or “Transcribing on CPU…”. Leave the tab open until the transcript appears.
- Save with “.srt”, “.vtt”, “.json” or “.txt”Each file is named after the one you chose. Corrections typed into the Plain text box go into Copy and the .txt file; the subtitle and JSON files are built from the timed segments as Whisper produced them.
FAQ
Is my audio uploaded when I transcribe it?
No. Decoding and speech recognition both run in your browser tab. The page downloads the Whisper model from Hugging Face and the ONNX Runtime WebAssembly engine from the jsDelivr CDN, and neither request contains your audio. A recording made with “Record” is kept in the page as a file and transcribed the same way.
How big is the model download, and does it happen every time?
The note under the model gives the download for the device it will run on. On the CPU that is about 41 MB for Speed, 77 MB for Balanced and 249 MB for Accurate. When your browser supports WebGPU the tool loads 16-bit weights instead (76, 146 and 485 MB), or 32-bit if the GPU lacks half-precision support (152, 291 and 968 MB). A 22 MB runtime file is fetched as well. The model files are stored in your browser's cache, so later runs in the same browser skip that download.
Which languages can it transcribe?
Speed transcribes English only. Balanced and Accurate take the language you choose from the Language list of 36: English, Arabic, Spanish, French, German, Hindi, Urdu, Chinese, Japanese, Russian, Portuguese, Italian, Dutch, Turkish, Korean, Polish, Ukrainian, Persian, Indonesian, Malay, Vietnamese, Thai, Tamil, Bengali, Swahili, Greek, Hebrew, Czech, Romanian, Hungarian, Swedish, Norwegian, Danish, Finnish, Filipino and Catalan. On Auto-detect, the default, Whisper listens to the first 30 seconds and picks from all 99 languages it knows, and the result names the language it detected. Choosing the language yourself is the surer option when you know it.
Can it make SRT or VTT subtitles from a video?
Yes. Choose the video file and keep Timestamps on Segments: your browser decodes the sound track and the tool resamples it to 16 kHz mono for Whisper. After transcription, “.srt” and “.vtt” save subtitle files with a start and end time for each segment. The line-length setting under Advanced (42 by default) moves text beyond that length onto a second line, and 0 keeps each caption on one line.
Can it translate speech into English?
Yes, with Balanced or Accurate. Choose the spoken language, set Task to “Translate → EN”, and Whisper writes the transcript in English. Translation into other languages is not offered, and the Task buttons are disabled on Speed because that model is English only.
Is there a limit on recording length or file size?
The tool sets no maximum duration or file size. It decodes the whole file in memory and then transcribes it in 30-second windows with a 5-second overlap, so a long recording takes proportionally longer and needs more memory. Keep the tab open until the transcript appears.
Does it use my GPU?
If your browser exposes WebGPU and grants the page a graphics adapter, the model runs there and the progress line reads “Transcribing on GPU…”. Otherwise it runs on the CPU in WebAssembly with an 8-bit version of the model. The finished transcript is labelled GPU or CPU so you can see which was used.