Audio to Text (Transcription)

How to transcribe audio to text

1. Choose an audio or video file, or press Record from microphone and stop when you're done. MP3, M4A, WAV, WebM and MP4 all work, as do most other formats your browser can play.
2. Pick a model. Base is a good start; Small is more accurate, especially for languages other than English, but slower. The badge shows whether the model is already in your browser.
3. Pick the spoken language, or leave it on Detect automatically.
4. Press Transcribe. The first run downloads the model; after that, the text appears part by part, and you can cancel at any point and keep what's done.
5. Check the transcript (click a timestamp to hear that spot), then copy it or download it as TXT, SRT or VTT.

How it works, and what stays private

The transcription runs on Whisper, the speech recognition model OpenAI released as open source in 2022. This page runs it with Transformers.js and ONNX Runtime Web, so the whole job happens inside your browser tab: the file is decoded to audio on your computer, the model turns it into text on your computer, and the text stays in the tab until you copy or download it.

Being precise about the network matters here, so this is exactly what the page downloads:

• The model, from Hugging Face (huggingface.co). On first use the browser downloads the files of the model you picked. The browser then keeps them, so the next run doesn't download them again unless the browser clears its storage.
• The ONNX runtime, from jsDelivr (cdn.jsdelivr.net). The engine that runs the model is a WebAssembly file of about 21 MB, loaded the same way.
• Nothing of yours goes out. Your audio, video and transcript are never sent to Hugging Face, jsDelivr or EasyTest. As on the rest of the site, Google Analytics counts page views and that a transcription or download happened (with the model name, never the text), and Google AdSense shows ads outside the tool.

If your browser supports WebGPU, Whisper runs on the graphics chip, which is usually much faster. Otherwise it runs on the processor through WebAssembly (WASM), which works everywhere but is much slower. The tool shows which one your browser will use next to the model menu.

Which model should you choose?

All four are multilingual Whisper models. Bigger models make fewer mistakes but download more, need more memory and run slower. The download is larger on WebGPU because the GPU version uses a full-precision encoder.

Parameter counts are from OpenAI's Whisper repository, which describes the turbo model as an optimized version of large-v3 that transcribes faster with little loss in accuracy. How fast each one runs depends on your device far more than on anything else: a model that keeps pace with the recording on a laptop's GPU can take several times longer on a phone's processor. When a run finishes, the tool tells you how long it took on your device, which is the most honest speed guide there is.

How accurate is Whisper, and what does it get wrong?

Whisper was trained on 680,000 hours of audio collected from the web, which is why it copes with accents, background noise and technical vocabulary better than older systems. Its accuracy still varies a lot by language: the Whisper paper shows that languages with more training data get much better results, and English has by far the most. We don't quote an accuracy percentage, because the number depends on the recording more than on the model.

Mistakes to look out for:

• Names and rare words. People's names, company names, product names and acronyms are often written as a similar-sounding common word.
• Overlapping speakers. When people talk over each other, Whisper usually follows one voice and drops the other.
• No speaker labels. The transcript doesn't say who said what. Whisper has no speaker identification.
• Silence, music and noise. The Whisper paper notes that the model can get stuck repeating itself, skip the first or last words of a stretch of audio, or, rarely, write text that wasn't said at all. Long silent or musical parts are where this happens most.
• Numbers and punctuation. Figures may come out as words or digits, and sentence breaks don't always match the speaker's.

Tips for better results

• Record close to the speaker. A phone on the table in the middle of a meeting is far better than a laptop microphone across the room.
• Choose the spoken language instead of detection when you know it, particularly for short clips or recordings that start with music.
• Try Small before giving up on a recording: it fixes many of the mistakes Base and Tiny make, especially outside English.
• Cut long silences and music first with the Audio Cutter. It makes the job faster and avoids made-up lines in quiet parts.
• Keep the tab open and in front while a long file runs, and plug a laptop in.

Frequently asked questions

Is my recording uploaded anywhere?

No. The audio is decoded and transcribed inside this browser tab, and the transcript stays there until you copy or download it. The only downloads happen on first use: the Whisper model files from Hugging Face (huggingface.co) and the ONNX runtime from jsDelivr (cdn.jsdelivr.net). Nothing from your recording is sent to them.

Is there a length limit?

There's no sign-up and no per-day or per-minute limit. A single file can be up to 4 hours long. Long files are slow and the decoded audio takes memory (about 230 MB for an hour, plus the model), so the tool warns you about anything over 30 minutes.

Can it tell who is speaking?

No. Whisper writes down what is said but doesn't label speakers, and when two people talk at once it usually catches only the louder one. For interviews, turn timestamps on and add names yourself while you play the audio back.

Why is the first run slow?

The first run downloads the model you picked: 45 MB for Tiny without WebGPU, up to almost 3 GB for Large v3 Turbo on WebGPU. On WebGPU the graphics chip also has to be prepared. The browser keeps the model files, so later runs start much sooner.

Does it work on a phone?

In recent mobile browsers, yes, but phones have less memory and are slower. Use Tiny or Base and files under about 30 minutes. If the tab reloads or stops, the phone ran out of memory; a computer will handle the same file.

Sources

1. Radford et al. (2022), Robust Speech Recognition via Large-Scale Weak Supervision, arXiv:2212.04356 (the Whisper paper: training data, accuracy by language, failure modes)
2. OpenAI Whisper repository (model sizes and the turbo model)
3. Transformers.js documentation (running Whisper in the browser)
4. The ONNX Whisper models this page downloads, on Hugging Face (Base shown)

Your recording never leaves your device. The model runs in this browser tab; the only downloads are the model files from Hugging Face and the runtime from jsDelivr on first use, and nothing you record or transcribe is sent anywhere.