Transcribe audio to text with Whisper, right in your browser
This free audio to text converter turns recordings into editable text: interviews, lectures, podcasts, voice notes, meetings and videos you have made. Drop an MP3, WAV, M4A, OGG, FLAC or video file and get a timestamped transcript you can correct, search and download as plain text, SRT or VTT subtitles, JSON or a Word document.
It runs OpenAI’s Whisper speech-recognition model on your own device with transformers.js, so the recording is never uploaded — handy for confidential interviews, client calls or anything you would rather keep private. For live dictation into a text box while you talk, use Speech to Text; this page is for audio files you already have.
How to use the Audio to Text
- Pick a model: Tiny is quickest, Base is a good default and Small is the most accurate but slowest.
- Leave “Detect automatically” or choose the spoken language. Under “Write the text in”, pick “English” to translate other languages into English text.
- Drop an audio or video file, then press “Transcribe”. The first time, the model downloads with a progress bar; long files show each 30-second part and a live preview.
- Play the recording and follow along: click any time to jump there, and click the text to fix names or misheard words.
- Press TXT, SRT, VTT, JSON or Word to download, or “Copy text” to paste the transcript elsewhere.
Audio to Text features
Whisper on your device
OpenAI Whisper (MIT-licensed weights), converted to ONNX by Xenova (Apache-2.0), runs via transformers.js (Apache-2.0) and WebAssembly in a background worker.
Language detection
The tool listens to the loudest 30 seconds, picks the most likely of Whisper’s 99 languages and shows how sure it is.
Translate to English
Whisper can write an English translation of Hindi, Spanish, French and other speech in the same pass.
Timestamps and paragraphs
Every phrase keeps its start and end time, and a pause of 1.5 seconds or more starts a new paragraph.
Subtitle-ready files
SRT and VTT cues are split into at most two lines of about 42 characters, ready for video players and editors.
Long recordings
Files up to two hours are processed in overlapping 30-second windows, with progress, time left and a live preview.
Which Whisper model to choose
All three are 8-bit versions of OpenAI’s multilingual Whisper models. Bigger models make fewer mistakes, especially with accents, background noise and languages other than English, but take longer and need more memory.
| Model | Download | Speed we measured | Good for |
|---|---|---|---|
| Tiny | 41 MB | About 6× faster than real time | Quick drafts of clear speech |
| Base | 76 MB | About 2.7× faster than real time | Most interviews, lectures and podcasts |
| Small | 240 MB | Roughly real time; about 5 GB of memory | Accents, noisy rooms, non-English speech |
How the transcription works
The file is decoded in your browser, mixed down to mono and resampled to 16 kHz, the rate Whisper was trained on, using a windowed-sinc filter so high frequencies do not fold back into the speech band. Whisper then turns each 30-second window into an 80-band log-Mel spectrogram; its encoder builds a representation of the sound and its decoder writes text tokens along with time-stamp tokens. Windows overlap by five seconds on each side and the overlapping words are matched up, so nothing is repeated or dropped at the seams.
Speeds in the table were measured in Chrome on a mid-range laptop processor with one WebAssembly thread and a 55-second speech clip — at that pace Base handles a one-hour podcast in roughly 20–25 minutes. We also tried graphics-card (WebGPU) builds; they needed two to three times larger downloads and were not faster on typical laptop graphics, so the tool sticks to the processor.
Tips for an accurate transcript
- Accuracy depends mostly on the recording: one voice close to the microphone beats a far-away phone in a busy café. Reduce steady hiss first with the Audio Noise Remover.
- Cut long silences, music or off-topic parts with the Audio Cutter — Whisper occasionally invents a phrase such as “Thank you for watching” over silence or music.
- Names, technical words and numbers are the usual errors. Fix them in the transcript before exporting; the edits go into every file format.
- Speakers are not labelled. The text is split into paragraphs at pauses, not by who is talking.
- For song lyrics, isolate the voice first with the Vocal Extractor; singing is still harder than speech.
- To burn or attach the subtitles to a video, use Add Subtitles to Video; the Subtitle Converter turns SRT into other formats.
Audio to Text FAQ
How do I convert audio to text for free?
Drop your audio or video file on this page, keep the Base model, press “Transcribe” and wait for the progress bar. Then download TXT, SRT, VTT, JSON or Word, or copy the text. There is no sign-up and no minute limit.
Is my recording uploaded anywhere?
No. Decoding and speech recognition both run in your browser. Only the Whisper model and the transformers.js engine are downloaded, once, and your browser keeps them for next time.
Which languages can it transcribe?
Whisper was trained on 99 languages. Results are best in widely spoken languages such as English, Spanish, French, German and Portuguese; Hindi, Tamil, Bengali and other Indian languages work but with more mistakes, so use the Small model for them.
How long can the audio file be?
Up to two hours per file. Processing time depends on the model and your computer — on a typical laptop, Tiny needs about a sixth of the recording’s length, Base about a third and Small roughly the full length.
Can it tell different speakers apart?
No. Speaker identification is not included; the transcript is grouped into paragraphs wherever there is a pause.
Why are some words wrong or missing?
Background noise, music, overlapping voices, strong accents and rare names all cause errors. Try the Small model, choose the language instead of detecting it, and correct the remaining words by clicking on the text.
How is this different from Speech to Text?
Speech to Text types live while you speak, using your browser’s own recognition service, which may send audio to Google, Microsoft or Apple. Audio to Text works on existing files and keeps everything on your device.