How to convert audio to text
- Select an audio or video file (MP3, M4A, WAV, MP4 and more).
- Pick a model (Balanced is a good start) and the language spoken, or leave it on automatic.
- Click Transcribe. The first run downloads the speech model; text appears as it is recognised.
- Copy the transcript or download it as TXT, SRT or VTT subtitles.
Private AI transcription
Most transcription services upload your recording to their servers, and many keep it. This tool runs Whisper, the open speech-recognition model released by OpenAI, directly in your browser using Transformers.js. Your audio is processed on your device. Only the model files are downloaded from Hugging Face the first time, and your browser caches them for next time.
That makes it a good fit for confidential material: client calls, interviews, medical or legal dictation, therapy notes, or anything you’d rather not hand to a cloud service.
Choosing a model
| Model | Download | Speed | Accuracy |
|---|---|---|---|
| Fast (Whisper Tiny) | about 40 MB | About 5 min per 10 min of audio | Fine for clear speech |
| Balanced (Whisper Base) | about 80 MB | About 7 min per 10 min of audio | Good for most recordings |
| Accurate (Whisper Small) | about 250 MB | Several times slower | Best of the three |
The speeds are what we measured on an ordinary 12-thread laptop; phones and older computers are slower. You can keep working in another tab while it runs, and the text appears as it is recognised.
If your browser supports WebGPU, you can tick GPU acceleration. It can be much faster on a computer with a dedicated graphics card, but on laptops with built-in graphics it is often slower than the default, so try it on a short clip first.
Tips for better transcripts
- Choose the language spoken instead of automatic detection when you know it.
- Clear audio matters more than the model: reduce background noise and keep the microphone close.
- Expect mistakes with names, numbers and jargon, and check them before you rely on the text.
- Translate to English turns speech in other languages straight into English text.
Outputs
Copy the text, download a .txt file (with or without timestamps), or download SRT and VTT subtitle files for YouTube, video editors and players such as VLC.
Have a video? It works directly with MP4 and WebM, or extract the audio first with MP4 to MP3. Clean up the result with the AI Text Cleaner or count it with the Word Counter.
Frequently asked questions
How accurate is the transcription?
For clear speech in common languages, Whisper is very good, but it does make mistakes, especially with names, numbers, strong accents, crosstalk and background noise. Always check important details. The Accurate model makes fewer errors but is slower.
Is my audio really private?
Yes. The speech model runs inside your browser, so your recording is not sent to us or to an AI company. Only the model files are downloaded, once, from Hugging Face.
How long does it take?
It depends on your device. On an ordinary laptop we measured about 5 minutes per 10 minutes of audio with the Fast model and about 7 minutes with Balanced. Phones are slower. GPU acceleration can be much faster with a dedicated graphics card, but is often slower on built-in laptop graphics.
Can it make subtitles?
Yes. Download the SRT or VTT file and upload it to YouTube, or load it in a video editor or player such as VLC.