Audio to Text Converter

How automatic audio to text transcription works

When you upload a file here, it is sent to a speech recognition model that listens to the audio in small slices, predicts the most likely sequence of words for each slice, and stitches the predictions into running text. The model has been trained on many thousands of hours of transcribed speech, so it has a strong sense of which word sequences are probable in a given language. That is also why it occasionally writes a plausible phrase that is not what was said: it is guessing from sound and context, not understanding.

Punctuation is added by a second pass that looks at pauses, intonation and sentence structure. It is good at full stops and commas, less reliable with question marks and quotation. Language is detected automatically from the first few seconds of audio, which is why a file that starts with a long silence or music can occasionally be assigned the wrong language; trimming the beginning fixes it.

The whole process is fast because nothing is done by hand. A typical three-minute voice memo comes back in 10 to 30 seconds. That speed is the main advantage over human transcription, which costs money per minute and takes hours or days. The trade-off is accuracy on difficult audio, covered below.

Supported formats and the 25MB limit

The converter accepts the common audio and video containers you are likely to have: mp3, m4a (iPhone voice memos, Apple Music exports), wav, webm (browser recordings), mp4 and mov (video; the audio track is extracted), plus ogg, opus, flac and aac. If your file plays in a normal media player, it will almost certainly transcribe.

The limit is 25MB per file and five files per hour per visitor. What 25MB means in minutes depends on how the audio was encoded. As a rough guide: a 128kbps mp3 is about 1MB per minute, so 25MB is around 25 minutes; a 64kbps voice recording is twice that; an uncompressed wav at CD quality is about 10MB per minute, so you would get only two or three minutes. Video is the least efficient, since most of the bytes are pictures you do not need.

• Long wav file? Convert it to mp3 first with any free audio editor and it will shrink by around 90 percent with no audible loss for speech.
• Long video? Export the audio track only (most editors have an "export audio" option) before uploading.
• Recording over 30 to 40 minutes? Split it into parts. Each part counts as one of your five hourly uploads.

How accurate is audio to text, and what hurts it

On a clean recording of one person speaking clearly into a decent microphone, automatic transcription gets the large majority of words right; the mistakes tend to cluster around names, product terms, numbers and homophones. On harder audio the error rate rises quickly, and it is worth knowing which conditions cause the most trouble so you can either fix the recording or budget time for clean-up.

• Crosstalk. Two people speaking at once is the single biggest problem. The model has to pick one voice and usually mangles both. Meetings and lively interviews suffer most.
• Background noise. Traffic, cafe chatter, keyboard clatter and air conditioning all mask consonants, which are what distinguishes most words. Even a quiet hum lowers accuracy.
• Phone and conference audio. Phone lines cut off the high frequencies that carry s, f and th sounds, and video-call compression adds artefacts. Expect more errors than from a local recording.
• Accents and dialects. Models are trained mostly on standard varieties. Strong regional accents, non-native speakers and code-switching between languages all reduce accuracy.
• Distance from the microphone. A laptop mic across the table picks up room echo. A headset or phone held close to the mouth is dramatically better.
• Specialist vocabulary. Medical, legal and technical terms, company names and acronyms are often replaced with a common word that sounds similar.

How to clean up a transcript

Raw machine transcripts are usable for search and reference as they are, but if anyone else is going to read the text, give it a quick edit. The box above is editable, so you can do this in place and then copy or download the result.

• Read with the audio playing. Fixing from memory misses the errors that look plausible; hearing the original catches them.
• Fix names and terms first with find and replace. A misheard product name usually appears the same wrong way every time.
• Add paragraph breaks where the topic changes. Machine output comes back as one long block, which is hard to read.
• Decide on verbatim or clean. For a quote, keep the exact words. For notes, remove filler words, false starts and repetitions so the text reads naturally.
• Mark anything you cannot make out with [inaudible] rather than guessing.
• Add speaker names at the start of each turn if more than one person is talking.

What people use audio to text for

The same converter serves very different jobs, and what counts as a good transcript changes with the job.

• Lectures and classes. Students record a lecture, transcribe it and search the text for the part they missed. Accuracy on technical terms matters more than punctuation.
• Interviews. Journalists and researchers need quotable text. Verbatim accuracy matters, and crosstalk is the enemy, so a quiet room and one person speaking at a time pays off.
• Meetings. Notes from a call that nobody wrote down. The transcript is usually skimmed, then summarised, so a rough version is fine.
• Voice memos. Ideas dictated while walking or driving, turned into text for a document or an email. These are short, single-speaker and transcribe very well.
• Sales and practice calls. Reps transcribe their own calls to count how much they talked, which questions they asked and where the conversation stalled. The transcript is the raw material for a review.
• Speeches and presentations. Record a rehearsal, transcribe it, and you can see your actual word count, your filler words and how far you drifted from the script.

Privacy, limits, and when you need something bigger

Your file travels over an encrypted connection to our transcription provider, is transcribed, and is deleted; we do not keep the audio and we do not save the transcript on our servers. The text exists only in your browser until you copy or download it. If you close the tab, it is gone. For confidential recordings, that is the behaviour you want, but it also means you should download the transcript before leaving the page.

The free tool is deliberately small: 25MB per file, five files an hour, plain text output, no speaker labels, no timestamps and no storage. For a two-hour interview, a batch of recorded meetings, or anything you need to keep and come back to, you will need a full account or a dedicated transcription service. We would rather say that plainly than have you discover it on the third upload.

If what you really want to know is how you sound, a transcript is only step one. Pavone turns a recording of your practice call or presentation into a scored review of pace, filler words, clarity and structure, and lets you rehearse the same conversation against an AI partner until it lands. Start with the transcript here, then take it further when you are ready.