Audio to Text: Transcribe Audio Free and Privately
AIConvert audio to text in your browser with Whisper, the open speech-recognition model from OpenAI. Drop MP3, WAV, M4A or a voice memo and get a transcript you can edit, search and save as TXT, Word, SRT, VTT or JSON. The model runs on your own device, so recordings are never uploaded, and there is no minute cap, account or watermark.
- Runs on your device
- Whisper AI
- WebCodecs · GPU
- Works offline
Updated
How to transcribe audio to text
Step 1: Drop your recordings on the box above. MP3, WAV, M4A, FLAC, OGG and phone voice memos all work, and you can add several at once.
Step 2: Choose the Spoken language, or leave Detect automatically, and a Speech model. Each model card shows the exact download for your device.
Step 3: Press Transcribe and confirm the one-time model download. The text appears line by line while it works.
Step 4: Fix any word in the editor, click a timestamp to hear that line, then Copy the text or export TXT, DOCX, SRT, VTT or JSON.
Good to know
Settings explained
- Spoken language
- Picking the language skips detection and improves accuracy on short clips.
- Speech model
- Downloaded once after you confirm, then cached for offline use. Bigger models are more accurate but slower.
- Fastest: from 44 MB · good for clear speech
- Balanced: from 80 MB · recommended
- Accurate: from 252 MB · slower
- Best: from 566 MB · needs WebGPU
Formats
Batch-friendly: add many files and they are processed one after another — if one fails, the rest still finish.
How it runs
- Whisper speech model
- Speech recognition uses an open Whisper model. It downloads once after you confirm its size, then runs on your device — on the GPU via WebGPU when available.
- Hardware video engine
- Video is decoded and encoded by your device’s media hardware through WebCodecs — usually several times faster than real time.
Which speech model should you use?
Start with Balanced. It handles most interviews, lectures and voice memos well and downloads quickly. Move up to Accurate for strong accents, noisy rooms or quiet speakers, and use Fastest on phones and older laptops. Best gives the cleanest text but needs a desktop browser with WebGPU.
| Model card | No WebGPU | With WebGPU | Best for |
|---|---|---|---|
| Fastest (Tiny) | 44 MB | 65 MB or 123 MB | Clear speech; default on phones |
| Balanced (Base) | 80 MB | 113 MB or 209 MB | Most recordings; default on computers |
| Accurate (Small) | 252 MB | 325 MB or 589 MB | Accents, noise, several speakers |
| Best (Large v3 Turbo) | Needs WebGPU | 566 MB or 762 MB | Top accuracy on a desktop GPU; no translation |
What is audio transcription?
Audio transcription is turning recorded speech into written text, ideally with a timestamp for each line so you can find a quote again. Manual transcription takes several times the length of the recording; speech recognition does a first draft in minutes, and you correct it.
Here the recording is decoded to 16 kHz mono, the format Whisper expects, and the model reads it in windows of up to 30 seconds, cut at quiet moments so no word is split; silent stretches are skipped. It writes words with their timings, detects the language among the 99 Whisper was trained on, and adds punctuation. The editor then lets you fix names and jargon before you export.
How accurate is AI transcription?
With clear speech in a widely spoken language, Whisper gets most words right, and the mistakes cluster in names, numbers and specialist terms. Accuracy drops with background noise, music under the voice, people talking over each other and very short clips. A bigger model helps more than anything else you can change.
| Problem in the recording | What you notice | Fix |
|---|---|---|
| Fans, traffic, room hiss | Missed or invented words in quiet passages | Clean it with remove background noise first, or choose Accurate |
| Strong accent or fast speech | Similar-sounding words swapped | Choose Accurate or Best |
| Clip shorter than about 10 seconds | Wrong language detected | Pick the Spoken language instead of detection |
| Names, brands, technical terms | Plausible but wrong spellings | Search the transcript and correct them in the editor |
| Two people talking at once | Merged or dropped sentences | No model separates overlapping voices; fix by ear |
MP3 to text and other audio files
MP3, WAV, M4A, AAC, FLAC, OGG, Opus and WebM audio are decoded by the browser itself, and AIFF by a small reader built into the page. Rarer formats such as WMA, AMR and AC-3 go through an FFmpeg decoder that loads once. A video file works too, because only its sound is read.
The result opens in an editor rather than a plain text box. Search finds every mention of a word, a click on a timestamp plays that moment, and your corrections are kept in every export.
| Format | What you get | Use it for |
|---|---|---|
| TXT | Plain text, with or without [0:00] timestamps | Notes, quotes, pasting into an email or a chat |
| DOCX | A Word document, one paragraph per transcript line | Editing in Word, Google Docs or Pages |
| SRT | Numbered subtitle cues with start and end times | Uploading captions to YouTube or a video editor |
| VTT | WebVTT subtitles | HTML5 video players and course platforms |
| JSON | Every word with its start and end time | Developers, search indexes, custom tools |
Transcribe audio without uploading it
- Nothing leaves your device. Interviews, medical notes, meetings and voice memos are transcribed by your own browser. Upload services keep a copy on their servers, at least for a while.
- No minute quota. There is no monthly allowance, queue or sign-up, so a three-hour lecture costs the same as a ten-second clip: nothing.
- Batch. Add several recordings and each gets its own transcript tab.
- Translate to English. One switch writes the transcript in English whatever language is spoken (Fastest, Balanced and Accurate models).
- Works offline. Once a model is cached you can transcribe in airplane mode.
Convert audio to text on iPhone, Android, Mac or Windows
The page runs in any modern browser, so the steps are the same everywhere. What differs is where your recordings live and how much memory the device has.
iPhone and iPad
In Voice Memos, tap the recording, then … → Share → Save to Files. Open this page in Safari, tap the drop zone and pick the M4A from Files. Phones default to the Fastest model to save memory.
Android
Open the page in Chrome and pick the recording from your file manager or Downloads. Pixel phones’ Recorder app transcribes its own recordings; this page handles files from any app.
Mac
Drag files from Finder into Safari, Chrome or Edge. Chrome and Edge support WebGPU on Macs, which unlocks the Best model and speeds up the others.
Windows 10 and 11
Drag files from File Explorer into Chrome, Edge or Firefox. Windows voice typing (Windows + H) only types live speech; it cannot read a file, which is what this page is for.
Speech to text vs audio to text
Speech to text usually means live dictation: you talk and words appear, as with your phone keyboard’s microphone. Audio to text starts from a recording you already have. This page is the second kind. To dictate something long, record it first with the voice recorder, then drop the file here, so you also keep the audio.
Tips & troubleshooting
Video file? It works here, but video to text explains MP4, MOV and screen-recording specifics, and plays the picture beside the transcript.
Transcribing WhatsApp messages? WhatsApp audio to text opens .opus voice notes directly and copies plain text without timestamps.
Turn Timestamps in TXT, DOCX and copy off in the editor when you want clean paragraphs for a document.
Need captions on a video rather than a transcript? The subtitle generator times captions word by word and can burn them in.
Recorded a podcast or lecture? Run it through the YouTube chapters generator to get “00:00 Intro” timestamps from the same speech model.
Audio to Text (Transcription) FAQ
How do I transcribe audio to text for free?
Drop the recording on this page, pick the language and a speech model, and press Transcribe. The model downloads once with your consent and then runs on your device, so there is no fee, sign-up, minute limit or watermark. Correct any words in the editor and export TXT, DOCX, SRT, VTT or JSON.
How do I convert MP3 to text?
Drop the MP3 here and press Transcribe. MP3 is decoded by the browser itself, so there is nothing extra to install. Choose Balanced for clear speech or Accurate for noisy or accented recordings, then copy the transcript or download it as a text file or Word document.
How accurate is AI transcription?
Clear speech in a common language comes out mostly right; names, numbers and jargon need checking. Noise, music, overlapping voices and very short clips lower accuracy. Picking the language and a larger model are the two changes that help most, and every word stays editable before you export.
Can I transcribe WhatsApp voice messages to text?
Yes. Save the voice note from WhatsApp (it is an .opus file) and drop it here or on WhatsApp audio to text, which is set up for voice notes and copies plain text. Several notes can be transcribed in one go, and none of them leave your phone or computer.
How do I transcribe audio to text on iPhone?
Save the recording to the Files app (in Voice Memos: Share → Save to Files), open this page in Safari and pick it from the drop zone. The Fastest model is selected on phones because it needs the least memory; long recordings are easier on a computer.
Can it tell different speakers apart?
No. Whisper writes what was said, not who said it, and there is no speaker-labelling model for browsers under a suitable licence yet. In interviews, lines often break where the speaker changes, so adding names by hand is quick.
Is my audio uploaded, and what gets downloaded?
Your audio is never uploaded. The main download is the speech model itself, from the Hugging Face CDN, after you confirm its size; rare formats such as WMA or AMR also load an FFmpeg decoder from this site. Both are stored in your browser’s cache, so the next transcription starts straight away and works without a connection.
Is there a file size limit?
There is no upload limit because there is no upload. Long files are processed in chunks, so even hour-long recordings work on a desktop computer. There are no daily caps and no watermark.