Choose an audio file
EarScribe decodes MP3, WAV, M4A, OGG, FLAC and WebM in the browser and hands Whisper a 16 kHz mono copy.
Whisper online
Use OpenAI Whisper online for free, in your browser
EarScribe runs OpenAI's open-source Whisper models directly in your browser tab with Transformers.js and ONNX Runtime Web. Pick a model, choose an audio file, and get Whisper's timestamped transcript without installing Python or paying for an API.
- No API key, no Python, no install
- Tiny, Base, Small or Large v3 Turbo
- WebGPU when available, CPU otherwise
Drop an audio file or click to browse
MP3, WAV, M4A, OGG, FLAC, WebM — practical length depends on device memory
Review the transcript before you export
Using Whisper for free in the browser
Pick a Whisper model
Base is recommended for most recordings. Compare all four to trade download size and speed for accuracy.
Review Whisper's output
Every segment carries Whisper's timestamps. Replay, correct and export TXT, SRT, VTT or JSON.
whisper online
Which Whisper model to run, and what to expect
Whisper is a family of speech recognition models that OpenAI released as open source in 2022, trained on 680,000 hours of audio in many languages. Model size is the main trade-off: bigger models are more accurate and much heavier to run.
The four models in EarScribe
Tiny (39 million parameters) is the fastest and smallest download, useful on phones and for a quick draft. Base (74 million) is the recommended balance for clear speech. Small (244 million) copes better with accents and noise but needs more memory, ideally with WebGPU.
Large v3 Turbo (809 million parameters) is a faster version of Whisper's largest model with a much smaller decoder: 4 layers instead of 32. It is the most accurate option here and needs a capable device; about 6 GB of memory and WebGPU are recommended.
WebGPU or WebAssembly
When your browser supports WebGPU, the model runs on your graphics chip, which is much faster for the larger models. Otherwise EarScribe falls back to WebAssembly on the CPU: it works almost everywhere, just more slowly.
WebGPU is available in recent Chrome and Edge and is arriving in other browsers. If Small or Large v3 Turbo feels slow, the CPU path is usually the reason.
Where Whisper struggles
Whisper can produce plausible text during silence or music, repeat a phrase, or drop words when people talk over each other. It does not label speakers.
Timestamps come per segment rather than per word. Use them to jump back and check anything important before you publish.
Review the transcript before you export
Choosing a model
- 1
Phone or older laptop
Start with Tiny or Base.
- 2
Clear speech
Base is usually enough.
- 3
Accents or noise
Try Small on a device with 4 GB of memory or more.
- 4
Best accuracy
Large v3 Turbo, with WebGPU and 6 GB of memory or more.
What running Whisper in a browser can't do
- Speed depends entirely on your device; there is no server to offload long files to.
- Live streaming transcription, translation and speaker labels are not part of this tool.
Getting the most from Whisper
Whisper detects the language from the first 30 seconds, so start the file with speech rather than a long music intro.
Silence and music can make Whisper invent short phrases. Trim long silent stretches and check the start and end of the transcript.
Larger models help most with accents, noise and specialist vocabulary; for clear speech, Base is often enough.
Questions about this workflow
Is this the real OpenAI Whisper?
Yes. EarScribe uses OpenAI's Whisper weights converted to ONNX, published by the onnx-community organization on Hugging Face, and runs them with Transformers.js. The model is the same; only the runtime is different.
Which Whisper models are available?
Tiny, Base, Small and Large v3 Turbo. The model picker shows each download size and a memory guide before you start.
How is this different from the OpenAI API?
The API uploads your file to OpenAI, needs an API key and bills by audio length. Here the model runs on your device, so there is no key or fee, but speed depends on your hardware.
Why is the first run slow?
The browser downloads the model once and caches it. Later runs on the same device and browser start much faster.