Skip to content

Audio to text converter

Transcribe MP3, M4A, WAV, OGG and voice notes to text on your device with Whisper. Timestamped, editable, exportable. Nothing is uploaded.

Loading tool…

Drop a video or audio file hereMP4, MOV, WebM, MKV, MP3, M4A, WAV, OGG and more · up to 500.0 MB and 60 minutes on this device
  • Your audio never leaves this device: it is decoded and recognised inside this browser tab.
  • The first run downloads the speech model (44 MB fast, 81 MB more accurate, 247 MB best) and keeps it in your browser for next time.
  • Keep this tab open until it finishes. Long files take a while; speed depends on your computer.
  • Works best in Chrome or Edge on a desktop or laptop.

How it works

Turn a recording into text: interviews, lectures, podcasts, meetings, voice memos or WhatsApp voice notes. The tool runs OpenAI Whisper, an open speech-recognition model, inside your browser, so the recording is decoded and transcribed on your own device and never uploaded. Each line gets a timestamp; click it to hear that passage in the player, correct the wording in place, and download plain text, text with timestamps, a Word document, a PDF or subtitle files.

Most audio formats work: MP3, M4A/AAC, WAV, OGG and Opus voice notes, FLAC, AMR and WMA (formats your browser cannot decode are converted with the same FFmpeg engine the site's audio converter uses). Choose Fast (Whisper tiny, about 44 MB downloaded once), More accurate (Whisper base, about 81 MB) or, for Arabic, Best (Whisper small, about 247 MB). In our test on a single 2.1 GHz processor core, Fast got through English speech at about 4 times real speed and More accurate at about 2 times, so a 30-minute lecture takes roughly 8 or 15 minutes; Best ran at about 0.6 times real speed on Arabic. Your speed depends on your processor.

Accuracy depends on the recording. A close microphone and one speaker at a time give the best results; music, echo, several people talking at once and strong accents cause mistakes, so proofread names and numbers. Recordings up to 60 minutes work on a computer (Chrome or Edge recommended, keep the tab open) and up to 20 minutes on a phone.

Summary, action items and decisions

When the transcript is finished, a summary panel appears under it: a short summary, key points, action items and decisions, each with the time it was said (click it to play from there). Everything in it can be edited, and it downloads as Word, PDF or text next to the transcript and subtitle files.

It is a basic on-device summary: sentences are ranked by how central they are and picked from the transcript, and action items and decisions are sentences with phrases such as "we need to", "I will", "we decided", "يجب", "سوف" or "تم الاتفاق". It does not rewrite or invent anything, and it can be wrong, so read it through. A search box finds the passages that answer a question, with their time; the answers are passages from the transcript, not generated text.

How to use it

  1. Choose or drop an audio file such as MP3, M4A, WAV, OGG or an Opus voice note.
  2. Pick the language (or automatic detection) and the Fast, More accurate or Best model.
  3. Click Transcribe and wait for the lines to appear; the model downloads only the first time.
  4. Play any line from its timestamp, fix mistakes directly in the text and search for words.
  5. Download the transcript as Word, PDF, TXT, timestamped TXT, SRT or VTT, or copy it.

Frequently asked questions

Is the audio sent anywhere?

No. Decoding and speech recognition both run in your browser tab. The only thing downloaded is the speech model, once, from this site; it stays in your browser's storage for next time.

Can I transcribe WhatsApp voice notes?

Yes. Save the voice note (usually an .opus or .ogg file) and drop it in. Short notes finish in seconds after the model has loaded.

How accurate is the transcript?

Good on clear speech, weaker with background noise, music, overlapping speakers, accents and specialist vocabulary. The More accurate model makes fewer mistakes. Read the transcript through before relying on it, especially names and figures.

Which languages can it transcribe?

Whisper's 99 languages, including English and Arabic. Set the language yourself for the most reliable result; automatic detection listens to up to three stretches of speech (skipping silence and music) and shows what it found, with a one-click redo if it is wrong. It can also translate speech from those languages into English.

What is the maximum length?

60 minutes and 500 MB on a desktop or laptop, 20 minutes on phones and tablets. The limits exist because the whole recording and the model must fit in the browser's memory.

Why is the first run slower?

The speech model (about 44, 81 or 247 MB) downloads first. Your browser keeps it, so later transcriptions start almost immediately.

Related tools