About this tool
A transcript turns a voice note, a recorded lecture or a meeting recording into something you can search, quote and edit. Doing it by ear, pausing and rewinding every few seconds, is one of the more tedious tasks a phone can be asked to help with. This tool listens for you, on the same device the recording is already sitting on.
The trick is a small AI speech-recognition model called Whisper Tiny, which this page downloads once from this site’s own server and then runs entirely inside your browser using WebAssembly: the same sandboxed way a browser runs any other bundled code. Your recording is decoded to plain audio samples on your device and fed straight to that model in memory. It never becomes a network request. What does cross the network, the first time only, is the model itself: 40 to 75 megabytes of numbers that have nothing to do with your audio, cached by your browser so every later transcription on this device starts immediately.
Being small enough to run in a browser tab has a cost, and it is worth stating plainly rather than glossing over: Whisper Tiny is meaningfully less accurate than the full-size Whisper models or a paid transcription service, it has no idea who is speaking when several people talk, and background noise or a thick accent will show up as mistakes in the text. For a voice note, a solo lecture recording or a phone interview in reasonably quiet conditions, it does a genuinely useful job. For a boardroom recording full of crosstalk, treat the result as a rough first draft to correct by ear, not a final transcript.
Auto-detect is a fine default when the whole clip is in one language spoken clearly. Choosing the language yourself (particularly for Hindi and the other Indian languages this tool supports) tends to produce a noticeably better transcript, because the model spends its effort transcribing instead of also guessing what it is listening to.
How to use this tool
- Choose your recording. MP3, WAV, M4A, OGG or WebM audio, or an MP4, WebM or MOV video: only its audio track is used. Clips up to 30 minutes are supported.
- Pick the language and what you want. Auto-detect works for clear, single-language audio. Choosing the language directly (Hindi or another Indian language, for instance) usually helps. Choose Translate instead of Transcribe if you want an English version of speech in another language.
- Run it. The first run downloads a small speech-recognition model, so it takes a little longer. After that it is cached, and later transcriptions on this device start straight away.
- Read, copy or download. The transcript appears on the page with a word count. Download it as a .txt file, as an .srt subtitle file with timing, or copy the text directly.
What this tool does not do
- Clips are capped at 30 minutes. Trim a longer recording first.
- No speaker labels. Everything comes back as one transcript, however many people were talking.
- Accuracy trails a large cloud transcription service, especially with background noise, crosstalk or a strong accent the model was not trained heavily on.
- Needs an internet connection the first time, to download the model. After that it works offline on this device until the browser clears its cache.
- Only the audio track of a video is used. Nothing about the picture is read or kept.
Common questions
Does my recording get uploaded anywhere?
No. The audio is decoded and transcribed entirely inside your browser, using your own device's processor. Nothing about the recording (not the file, not the transcript) is sent to a server. The only thing that travels over the network is the speech-recognition model itself, downloaded once from this site and cached by your browser after that. The model is hosted here, so your browser never contacts a third party for it.
Why does the first run take longer?
It is downloading the model: Whisper Tiny, a small AI speech-recognition model from OpenAI, roughly 40 to 75 MB depending on how your browser caches it. That happens once. Every run after that, on this device, starts straight away because the model is already there.
How accurate is it?
Whisper Tiny is the smallest version of a well-regarded speech model, chosen here because it is realistic to run inside a browser tab. It does a solid job on clear speech in a fairly quiet room. It is noticeably less accurate than the larger Whisper models or a paid cloud transcription service, and it struggles more with background noise, overlapping speakers, and heavy or unfamiliar accents.
Does it work for Hindi and other Indian languages?
Yes, for the languages listed in the dropdown, because the underlying model was trained on them. Accuracy in Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi and Urdu is generally behind its English accuracy, and picking the language directly (rather than leaving it on auto-detect) usually helps.
What does Translate do?
It produces an English transcript of speech in another supported language, using the model's own translation mode rather than transcribing and then translating separately. It only translates into English, not between other language pairs.
Can it tell me who said what?
No. It produces one continuous transcript with no speaker labels, so an interview or a multi-person meeting will read as a single unbroken block of text regardless of how many people were speaking.
Can I use it for a live call or a meeting as it happens?
No. It works on a file you already have, not on a live microphone feed, and it is not built for real-time captioning.