Split Audio for Transcription
Split a long recording into parts that fit the upload limit of your transcription service: 25 MB for Whisper and the OpenAI API, 10 MB for a synchronous Google request, or any size or length you choose. Smart split cuts at pauses, so no word is cut in half. No upload, no signup, no file size limit.
Split for transcriptionThe button opens the splitter with 25 MB per part already set. Background: the right chunk size for every transcription API.
Parts that fit the limit, in 3 steps
A two-hour interview becomes numbered parts in well under a minute. Nothing installed, nothing uploaded.
Drop the recording
MP3, M4A, WAV, FLAC and most other formats, hours long if need be. The file stays on your device.
Set the size per part
25 MB is already selected, the OpenAI limit. Type another size, or a length in minutes. The line under the field says how many parts you will get.
Split and download
Smart split moves each cut to the nearest pause; Quick split cuts at the exact size. Download the parts one by one or all of them as one ZIP, numbered in order.
What the size counts for
The size applies to the file you download, because that is what you upload next. Free downloads are WAV, which holds about 10 MB per minute, so 25 MB is about two and a half minutes of audio per part: fine for a short interview, a lot of files for a two-hour lecture. With Pro the parts stay MP3 or M4A at the original quality, which fits roughly ten times as much audio into the same size: 25 MB of 128 kbps MP3 is about 27 minutes. For a long recording that is the difference between five parts and fifty. The line under the length field shows the minutes per part for the format you will download, so there is no guessing.
If you would rather split by time, type minutes instead: 15 minutes of 128 kbps MP3 is about 14 MB, safely under every limit in the table below, and short enough that a failed upload costs little.
The limits of the common services
As of October 2026. The guide keeps the full table and the reasoning.
OpenAI: Whisper and gpt-4o-transcribe
25 MB per request, and 25 minutes of audio for gpt-4o-transcribe. Use 25 MB per part, or 15 minutes.
Google Speech-to-Text v2
10 MB and 1 minute for a synchronous request, up to 8 hours per file in batch. Use 1 minute for sync, 5 to 10 minutes for batch.
Amazon Transcribe
2 GB and 4 hours per batch job. The limit is rarely the problem; 15 to 30 minute parts keep retries cheap.
Azure Speech
300 MB and 2 hours for fast transcription, 1 GB in batch. 15 to 30 minute parts.
Google Gemini
20 MB inline in a prompt, larger files through the Files API, up to 9.5 hours of audio. 20 MB per part for inline use.
AssemblyAI, Deepgram, ElevenLabs
2 to 5 GB and up to 10 hours. Here the parts are for your own workflow: 30 to 60 minutes keeps a failed request from costing an hour.
Why the cut should land on a pause
Each part is transcribed on its own. A word cut in half comes out garbled or missing at the end of one part and the start of the next, and the next sentence loses its context. Smart split listens around each planned cut and moves it to the nearest pause, which costs a few seconds of part length and no words. For music, or when your pipeline overlaps the parts and merges the text itself, Quick split cuts at the exact size.
Every part keeps the sample rate and channel count of the original, and with Pro the audio inside is the original MP3 or M4A, untouched. Nothing is re-encoded on the way to the transcription service.
Putting the transcripts back together
The parts are numbered in order. The chapter marks list under the results gives the start time of every part, for example 00:00, 14:52, 29:40; add that offset to the timestamps each transcript comes back with and the whole thing lines up with the original recording. Copy the list into your notes, or download it as a text file.
What people split for transcription
The recordings that arrive here most often.
Interviews and research
Hour-long qualitative interviews, split at pauses, transcribed part by part, stitched with the start times.
Meetings and calls
Recorded Teams, Zoom and phone calls that are too long for one upload. See also splitting recorded meetings.
Lectures and courses
Two-hour recordings into 15-minute parts, transcribed and summarised one by one.
Podcasts and show notes
Episodes transcribed for show notes and chapters; the marks list doubles as the chapter list.
Voice memos
Long dictations from the phone, split into parts a transcription app accepts.
Legal and medical dictation
Recordings that must not go through a third-party converter: the split happens on your own machine.
Frequently asked questions
What is the file size limit for Whisper?
The OpenAI transcription endpoint accepts 25 MB per request, for whisper-1 as well as gpt-4o-transcribe; gpt-4o-transcribe also stops at 25 minutes of audio. A longer recording has to be split first, which is what this page is for.
How many minutes fit into 25 MB?
That depends on the format of the parts. 25 MB holds about 27 minutes of 128 kbps MP3, about 10 minutes of 320 kbps MP3, and only about two and a half minutes of 16-bit stereo WAV at 44.1 kHz. Free downloads are WAV; with Pro the parts stay MP3 or M4A at the original quality, which fits roughly ten times as much audio into the same size.
Does splitting change the audio?
No. MP3, M4A, AAC and WAV files are split by copying, without decoding the whole file, so with Pro each part is the original audio untouched. The free WAV download is the same part decoded once, nothing re-compressed. Every part keeps the sample rate and channels of the original, which transcription services expect.
Will words be cut in half?
Not with Smart split. It listens around each planned cut and moves it to the nearest pause, so every part begins and ends between words. Quick split cuts at the exact size instead, which is fine for music or when the service merges overlapping text itself.
Is my recording uploaded anywhere?
No. ChunkAudio runs entirely in your browser. The recording is read and split on your own device, which matters for interviews, patient notes, legal dictation and anything else you would not hand to a converter site. The parts go to your transcription service from your computer, as before.
Can I split a three-hour recording on a phone?
Yes, for MP3, M4A, WAV and Ogg (Opus or Vorbis, the voice-note formats). Those are cut by copying, without decoding the whole file, so the length does not matter. FLAC and other formats have to be decoded first, which tops out at roughly two hours on a computer and less on a phone; convert those to MP3 or M4A first.
Which format should I send to the transcription service?
MP3 at 128 kbps or an M4A: they hold the most speech per megabyte and every service accepts them. WAV wastes most of a 25 MB limit on silence and headroom. If your recording is already MP3 or M4A, the Pro download keeps it that way; otherwise download WAV and convert, or split by minutes instead of size.
How do I keep the timestamps right across the parts?
After the split, the chapter marks list under the parts gives the start time of every part, for example 00:00, 14:52, 29:40. Add that offset to the timestamps the service returns for each part and the transcript lines up with the original recording. The parts are numbered, so the order survives too.
Is it free?
Yes. Splitting is free and unlimited, with no signup and no watermark. The optional Pro licence (a 7-day pass from €4.99, a yearly plan or a lifetime licence) keeps the parts in the original MP3 or M4A format, splits several files in one go, saves presets and exports the marks as a CUE sheet or chapters file, and removes ads.
Parts that fit, now
Drop the recording, keep 25 MB or set your own size, split. The recording never leaves your browser.
Split for transcription