Bottom Line Up Front (BLUF): The tool turns audio or video into text and subtitles: it recognises speech with the Whisper model right in your browser, detects the language of the recording or takes the one you pick, adds timestamps and gives you the result as TXT, SRT, VTT and JSON. For video it also builds a clip with burned-in subtitles. The file is never uploaded, no tokens are charged and there is no paid API here.
The tool takes audio or video and turns speech into text. You upload a file and get a transcript with timestamps, ready-made subtitles and, for video, the same clip with subtitles burned right into the frames.
Recognition runs on the Whisper model (the open multilingual Xenova/whisper-tiny model) through Transformers.js: on the graphics card via WebGPU when the browser offers it, otherwise on the processor via WASM. The audio track is extracted from the file locally, in the memory of your browser.
Audio: MP3, WAV, M4A, AAC, OGG, Opus, FLAC. Video: MP4, WebM, MOV, M4V, MKV, AVI, OGV. For video the tool pulls the audio track out by itself — you do not need to extract it separately.
The sound is converted to mono 16 kHz — exactly the form Whisper expects. A long recording is recognised in 30-second chunks with a 5-second overlap, so words are not lost at the joins.
The file is not uploaded: the audio, the recognition and the subtitle assembly all happen inside your browser. There is no paid API here and no tokens are charged.
The model weights are downloaded once from the open Hugging Face catalogue and then stay in your browser cache, so the next visit loads the model without downloading it again. There is no server-side inference, no CUDA and no fallback GPU on our side.
Whisper recognises 99 languages. By default the language is detected automatically and the tool shows which language it heard. If the recording is short or mixed, you can pick the language manually from the list — recognition is then more accurate.
You can copy the text with one button or download it in four formats. TXT is a plain transcript, with or without timestamps. SRT and VTT are ready-made subtitles for players, video editors and YouTube. JSON is the transcript with the language, the duration and every line with its timestamps — handy for further processing.
For a video file the tool builds a new clip in which the subtitles are burned right into the frames. Such a video plays anywhere, even where external subtitles are not supported: in messengers, on social networks and in any player. The original sound is kept.
The same subtitles are also available as SRT and VTT files, so you can attach them to the original video without rebuilding it.
It does not translate speech into another language: recognition runs in the language of the recording, translation is a separate task. It does not separate speakers by voice and does not label who said what.
It does not treat music and noise as speech: if there is no intelligible speech in the recording, the transcript will be empty or inaccurate. Quality depends on the recording — clean speech without noise and echo gives a noticeably better result.
Choose an audio or video file, set the language if you want (or leave automatic detection) and press “Recognize speech”. Wait for the transcript — the progress is shown on screen and recognition can be cancelled.
Then copy the text or download TXT, SRT, VTT or JSON. If you uploaded a video, press “Make video with subtitles” and download the finished clip.
Video and podcast creators — subtitles for publishing and a transcript for descriptions. Students and journalists — text from lectures, interviews and voice recordings. Anyone who needs text from a recording quickly without sending it to someone else's server.
Nowhere. The audio is extracted and recognised entirely inside your browser on your device. The file never reaches our server.
No. This is a free tool that runs on your device: there is no paid API here, so there is nothing to charge.
The model computes on your device. The first run takes longer: the model weights are downloaded and the graph is warmed up. After that the model comes from your browser cache and recognition is faster.
Pick the language manually in the list and run recognition again — on short recordings a manual choice is more accurate than automatic detection.
Yes. Download SRT or VTT and attach them to the video, or build a clip with burned-in subtitles right in the tool.