How to Transcribe Audio with AI
MAI Transcribe 2 · MAI Transcribe 2
A working method for turning a meeting or interview file into text you can edit, including formats, failure cases, and when to stop.
Transcribing audio with AI is a three-step job: give the model a file it can hear, run one pass, then clean up names. This page describes the workflow used on maitranscribe2.com. It is an independent guide, not product documentation.
If you only need a labeled draft of a meeting or interview, start on the homepage upload. For scene-specific notes (captions, podcasts, clinical dictation), use the speech to text page.
What you need before you upload
Export a real file. Screen-record rambles and WhatsApp voice notes work; a 90-minute dump of mixed rooms usually does not, until you split it.
Formats this site accepts
MP3, WAV, M4A, FLAC, AAC, OGG, and MP4 / MOV / WebM video (audio track only). After signing in, each account gets one successful free job per UTC day, up to 5 minutes and 10 MB. Plus allows 30 minutes; Pro allows 60 minutes.
A file that is “clean enough”
You do not need a studio. You do need:
- Speech louder than the music bed
- One language most of the time (switching is fine; three languages in one sentence is a gamble)
- A clip you can actually review. Twelve minutes of a two-person interview is a better first test than a three-hour board meeting.
If your editor can strip leading silence, do that. Do not over-normalize until the room noise becomes a hiss the model treats as words.
Run the first pass
Open the homepage tool. Leave the mode on Verbatim for the free try. Verbatim keeps ums and repeats; that is useful when you are checking whether the model heard the file at all. Paid plans can switch to Smart for a cleaner reading copy, and can attach a phrase list (names, drug names, product names) of up to 100 terms.
What a usable first pass looks like
On a 12-minute interview we ran internally (two speakers, a café, English with a few Mandarin names):
- Speakers were labeled on most turns
- Timestamps were close enough to jump in a player
- Two proper names were wrong; one laugh over the espresso machine became a phantom word
That is a pass you can edit. A pass that invents a third speaker or drops a whole answer is not. Re-export a louder file before you buy more minutes.
Clean and publish
Copy the transcript. Fix names against a guest list, not against your memory. For captions, keep timestamps and export SRT or VTT on a paid plan. For notes, delete filler and add headings.
Failure cases we see on this site
- File rejected: wrong extension or over the plan size/duration. Re-export M4A or MP3; split the recording.
- Empty transcript: the model returned no text. Credits are not taken on a failed paid job. Try WAV, or a shorter slice from the middle of the file where people are actually talking.
- One speaker for everyone: common on a single microphone in a loud room. The draft is still usable as a document; do not trust it as a “who promised what” record.
- Wrong language mixed in: usually a music bed or a TV in the background. Trim the intro.
When this method is the wrong tool
If the audio cannot leave the machine, run a local model instead. Cost and ops for that path are in What it costs to run Whisper yourself. Accuracy tables versus Whisper live on MAI Transcribe 2 vs Whisper, not here.
Test conditions for this article
- Site: maitranscribe2.com, browser upload, September 2026
- Free path: sign in for one file per account per UTC day, Verbatim only, no phrase list
- Paid path: credits at 1 credit ≈ 1 minute, speaker labels and SRT available
- We do not claim medical or legal accuracy. Read a line before you quote it.
Operator and contact: About.