Local Audio Transcriber
Transcribe speech from an audio file to text using Whisper, running entirely on your device via WebAssembly. Supports 99 languages (pick the one spoken from a dropdown — this can't auto-detect it) and handles recordings of any length by processing them in overlapping chunks, not just short clips. Built for clear, single-speaker audio: a voice memo, a recorded conversation, a podcast clip. This is genuinely the largest download and slowest tool on the site: only the full-precision model weights (~150MB) actually load reliably, since the smaller quantized versions hit a real WebAssembly runtime compatibility issue, and CPU-only transcription takes real time for longer recordings. Nothing about your audio is ever uploaded — the model downloads once and every file after that transcribes completely offline.
The model can't detect this on its own — pick whatever language the audio is actually in, or it'll try to force everything into English and produce garbled text.
About this tool
This tool transcribes speech from an audio file to text using Whisper, a real speech recognition model, running entirely on your device via WebAssembly. It supports 99 languages, selected up front from a picker — this library's version of Whisper has no language-identification step of its own, so rather than silently assuming English (and mangling anything else into garbled, repeating English-sounding text trying to force-fit foreign phonetics into it) it asks. It also processes long recordings in overlapping chunks rather than just the first 30 seconds, so it holds up on anything from a short voice memo to a lengthy meeting recording. It's genuinely useful for a private voice memo, a recorded conversation, or any audio you don't want sent to a transcription service just to get the words out of it. Being upfront about the real cost of that: this is by far the largest download and the slowest tool on the entire site — the model itself is roughly 150MB (a compatibility issue with quantized, smaller versions of this model meant only the full-precision weights actually work reliably, so there was no smaller option to fall back to), and transcription runs on your device's CPU with no GPU acceleration, so a several-minute recording will take real time to process — a progress indicator shows how far through the audio it currently is. Clear, single-speaker speech works best; heavy background noise, overlapping speakers, strong accents, and non-English languages (this is the smallest Whisper size, and it's noticeably weaker outside English) will all reduce accuracy. Nothing about your audio is ever uploaded — the model downloads once and every file after that transcribes completely offline.
What you can do
- Transcribe speech from an audio file using a real on-device speech recognition model
- Supports 99 languages — pick the one actually spoken, not just English
- Copy the transcript or download it as a .txt file
- Word count shown alongside the result
- Your audio is never uploaded — transcription runs entirely in your browser
How to use
Step 1
Pick the spoken language
The model can't detect this on its own — select it up front, or it'll assume English.
Step 2
Upload an audio file
A voice memo, a recorded meeting snippet, a podcast clip — clear speech works best.
Step 3
The model transcribes it
Runs entirely on your device via WebAssembly, in chunks so it holds up on long recordings — this is the slowest tool on the site, so it can take a while.
Step 4
Copy or download
Get the transcript as plain text, with a word count.
Example
A short voice memo:
Result
A plain-text transcript, with a word count shown alongside it.
Clear, single-speaker audio with minimal background noise transcribes noticeably better than a noisy recording or a crowded conversation with overlapping speakers.
How your data is handled
This tool is labeled 🔵 Local AI: your audio is read directly using your browser's File API and transcription runs entirely on your device via a local WebAssembly model. It's never uploaded anywhere. The one network request this tool makes is downloading the model files (~150MB total) once, which are cached by your browser afterward — every audio file you process after that runs completely offline.
Frequently asked questions
Why is the model download so much bigger than this site's other local AI tools?+
The smaller, quantized versions of this speech model failed to load at all due to a real compatibility issue in the underlying WebAssembly runtime — not a choice to skip optimizing it. Only the full-precision weights (~150MB) actually work reliably, so that's what's used, clearly disclosed rather than glossed over.
Why is it slow?+
Speech recognition is far more computationally demanding than the site's other on-device models, and this runs on your CPU via WebAssembly with no GPU acceleration. A short clip transcribes in seconds; a longer recording will take real time — there's no way around that for a model this size running entirely client-side.
What languages does it support?+
All 99 languages Whisper was trained on — pick the one actually spoken from the dropdown before uploading. It can't auto-detect the language itself, so getting this right matters: fed the wrong language, it doesn't fail cleanly, it tries to force what it hears into that language's words, which reads as garbled, repetitive nonsense. Accuracy itself is generally strongest for English and noticeably rougher elsewhere, since this is Whisper's smallest, weakest size.
How does it handle long recordings?+
It processes audio in overlapping ~30-second chunks rather than only looking at the first 30 seconds, so a multi-minute or multi-hour recording gets transcribed in full rather than degrading into repeated text partway through. A progress indicator shows how far into the audio the transcription currently is.
Does it handle multiple speakers or background noise well?+
Not especially — it works best on clear, single-speaker audio. Overlapping speakers, heavy background noise, and strong accents will all noticeably reduce accuracy.
Is my audio uploaded anywhere?+
No — it's read directly in your browser and transcribed on your device using a local WebAssembly model. It never leaves your device, and nothing is retained once you leave the page. The only network request is a one-time model download (~150MB), cached by your browser afterward.