How to convert speech to text online and translate it
- Choose a model size (Base by default). The first use downloads the model once and caches it in your browser.
- Press the big microphone button and speak; pause briefly after each sentence and the text appears. You can also upload MP3, WAV, M4A and other recordings.
- Edit the transcript directly, copy it in one click or download a TXT file; it is saved automatically to this browser’s history.
- Tick target languages and click Translate to get several translations at once, or open the text in the full translator.
Model sizes and when to use them
| Model | First download | Best for |
|---|---|---|
| Whisper Tiny | 42 MB | Fastest · short phrases |
| Whisper Base | 76 MB | Recommended · balanced |
| Whisper Small | 240 MB | Most accurate · needs a strong PC |
Frequently asked questions
Is my recording uploaded to a server?
No. Microphone audio and uploaded files are decoded inside your browser tab and transcribed locally by the Whisper model. On first use your browser downloads a pinned version of the model files from Hugging Face (model files only – no audio is sent) and then loads them from the browser cache.
What is the difference between the three models?
Tiny is about 42 MB and the fastest, good for short phrases; Base is about 76 MB and the recommended default; Small is about 240 MB and the most accurate, but needs a capable computer or a WebGPU browser. Each model downloads once and the cache can be deleted from the page.
Which languages are supported? Is the language detected automatically?
The multilingual Whisper models cover about 99 languages including English, Mandarin, Cantonese, Spanish, French, German, Japanese, Korean, Russian and Arabic. The spoken language is detected automatically; choose it manually for strong accents or mixed-language speech.
Why does text appear after I pause?
To run reliably in a browser, each sentence is sent to the model when you pause, so text appears sentence by sentence – usually 1–3 seconds after the pause. Continuous speech is also split automatically about every 24 seconds.
Is the translation done on my device?
If your Chrome has the built-in Translator API (desktop Chrome 138+), translation happens on your device. Otherwise, after you click Translate, only the text (never audio) is sent to XunL and translated by Google Cloud Translation. Untick “Allow the XunL translation service” to turn that path off.
Which browsers work?
Desktop Chrome or Edge are recommended (WebGPU makes it fast). Firefox, Safari and mobile browsers run Whisper on WebAssembly (CPU) – it works but is slower, so choose Tiny or Base. Microphone permission is required; on iPhone use Safari and keep the page in the foreground.
How accurate is it? Can I use it for meeting notes?
With clear speech in a quiet room, Base and Small are usually good enough to use directly for English and Mandarin, but names, model numbers and figures can be wrong. The tool is meant for quick notes, dictated drafts and on-the-spot translation; check important documents by hand.
Where is my history stored?
Only in this browser’s localStorage (up to 30 entries); nothing is synced to a server. Clear your browser data or use “Clear history” to remove it.