Audio File Too Large for Transcription? Shrink It Below the Cap

The Whisper 25 MB limit is the cap in OpenAI’s speech-to-text guide, which says files “can be up to 25 MB”, in mp3, mp4, mpeg, mpga, m4a, wav or webm. Length does not matter on its own; bytes do. Audio size is bitrate times duration, so a recording that is long, high-bitrate or uncompressed crosses it fast.

The page’s own advice for longer recordings is to use a compressed format or split the file, and to avoid splitting in the middle of a sentence. Compressing first usually means no splitting at all, which keeps one transcript and one set of timestamps. Other services set different caps, as the table shows.

At 64 kbps, about 52 minutes fit in 25 MB; at 32 kbps, about 104. The formula is bitrate in kbps times 1,000, times seconds, divided by 8, so the table below is arithmetic, not a promise about any one file. An encoder’s delivered bitrate can drift from the requested one, so always read the size of the result.

Leave a margin. A file at 24.8 MB is a gamble if the service counts megabytes differently than your Finder does, so aim at 20 MB for a 25 MB cap. That is a habit, not a rule from the provider.

Otter publishes its import limits in its own help center, so check the current numbers there before a batch. We do not quote a figure here.

What carries over from the services above: a smaller file uploads faster and fails less often, and compressing to a voice bitrate is the same job whichever service receives it. If your upload comes from a video recording, export the audio on its own first and re-encode that. Zoom users can start with the Zoom recordings page.

Choose Compress, set the output to M4A (or MP3), pick a bitrate from the table, and switch the channels to mono. Smol’s audio bitrate runs from 16 to 512 kbps and its sample rate from 8 to 192 kHz. M4A and MP3 are both on OpenAI’s format list; OGG and FLAC are not, so do not choose them for that API.

Sample rate barely moves a lossy file at a fixed bitrate: in audio bitrate explained five sample rates of MP3 at 48 kbps (encoded with ffmpeg) differed by 216 bytes. Groq’s docs say it downsamples to 16 kHz mono before transcribing, so there is no need to chase the sample rate for size. We make no claim about how bitrate changes the transcript, so spot-check one result before you convert a hundred.

Drop the recordings into Smol, set M4A, 64 kbps, mono, run the batch, then sort the outputs by size and upload the copies. Originals stay untouched by default, so you keep the full-quality master.

1. 1

   Work out your bitrate

   Take the cap (25 MB, or about 20 for a margin) and the longest recording. A 90-minute interview at 64 kbps is about 43 MB, so it needs 32 kbps (about 22 MB) or a different service.

2. 2

   Drop the files or the whole folder in

   Smol takes mixed formats in one batch, such as WAV, M4A, MP3 and FLAC.

3. 3

   Set Compress, M4A, your bitrate and mono

   Save it as a preset. Presets are per Mac, so write the numbers down for a second computer.

4. 4

   Run it and check the size

   Output is a copy with a _compressed suffix beside the original by default. Sort by size. Anything over 20 MB goes again at a lower bitrate.

5. 5

   Listen to one before you upload all

   Play a minute from the middle of the quietest speaker. If it is unclear to you, raise the bitrate.

Yes: record in a compressed format to begin with. Apple’s Voice Memos has an Audio Quality setting that offers Compressed (lower-quality, smaller files) or Lossless (high-quality, larger files). Compressed is the smaller choice for interviews. Zoom, Teams and phone recorders have their own settings; check which one your recorder uses.

The tradeoff is that you cannot get detail back later. If the interview is the only copy of something you cannot repeat, keep the lossless original and compress a copy for transcription. Compressing a file that is already a low-bitrate MP3 gains little, as the compress audio guide explains.

Free options come first, and several are enough for one file a week. Smol earns its place when there is a folder of them.

Confidentiality is a real difference. Interview audio can carry names and consent conditions; an online converter receives your file on its own server, while Smol does not upload anything. The confidential documents page covers the same reasoning. Podcast editors have a parallel problem, covered on the podcasters page, and a long call can come from a webinar, handled on the webinar recordings page.

Can an AI agent run the batch?

Yes, through Smol’s MCP server, described on the MCP page, so a coding agent can run a folder of jobs locally. A one-off needs only the app.

You need the transcript itself. Smol does not transcribe. It makes the file smaller; the service you upload to makes the text.

The file is already small. A 32 kbps MP3 gains little from another pass, and Smol keeps the original if compressing would make a file larger.

You need to clean up noisy audio. Smol is an encoder, not an editor: no loudness normalization, no noise reduction, no trimming. Use your audio editor, then Smol.

You are not on a Mac. Smol runs on macOS only. A Windows laptop in the field needs another tool.

You transcribe one clip a month. Terminal afconvert is free and already installed. Our audio compression guide shows the command.