How we measured

The audio

Six real meetings, ten minutes from each, one hour of speech in total. All of them are in English. The audio is the file Aside itself records and stores, compressed AAC, so every engine was given exactly what the app would have given it. No studio recordings and no read-aloud scripts.

The reference

A transcript corrected by hand, word by word. Where the systems disagreed with each other, we listened to that passage and decided it individually rather than letting any one engine cast the deciding vote. An earlier version of this run scored against a reference written by a model, and it produced a different and flattering answer: a model that shares the same habit of dropping quiet speech will not penalise that habit. Every number on this page is scored against the adjudicated reference instead.

What was compared

Four transcription systems on the same audio: AssemblyAI universal, Deepgram nova-3 and OpenAI’s whisper-1 through their APIs, and Aside’s own engine through a runner that reproduces the app’s batch pipeline exactly, on an M4 Pro. The hosted services were sent the same compressed file the app keeps rather than a cleaner copy of it, so nothing here is measured on audio the app would never have had.

How it was scored

Word error rate, micro-averaged: every error over the whole hour counted against every reference word over the whole hour, rather than averaging six per-meeting rates, so a short meeting cannot weigh as much as a long one. Scoring is jiwer with one written-down normalisation, applied identically to every engine and not tuned per engine.

Where two results sit close together, the difference is resampled to get a confidence interval around it, and it only counts as a difference when that interval clears zero. Several comparisons in this run did not clear it and are not stated here as differences.

Accuracy

Verbatim word error rate, so every filler word, false start and backchannel counts. Lower is better.

• AssemblyAI universal7.1 word error rate, 92.9 accuracy.
• Deepgram nova-37.4 word error rate, 92.6 accuracy.
• Aside on device9.9 word error rate, 90.1 accuracy.
• OpenAI whisper-111.4 word error rate, 88.6 accuracy.

The cloud services came out ahead of the on-device run. The gap between Aside and the best cloud result is 2.8 percentage points, and it is almost entirely words that were dropped rather than words that were heard wrong. Quiet, short interjections are where they go.

Speed

On an M4 Pro, Aside transcribes about 29 times faster than real time, which is roughly two minutes of work for an hour of audio. The cloud figure on the front page is a round trip and not a processing time: it counts uploading the audio, the service transcribing it, and the transcript coming back, because that is the wait a person actually has.

What this does not cover

Six meetings is a small sample, and one that is small enough that a different six could move the figures. Everything in it is English, so nothing here says anything about any other language. The audio is the compressed file the app stores rather than the uncompressed capture, which is the honest basis for a shipping app and not the most favourable one.

The numbers describe the transcript Aside writes when a meeting ends. The live transcript that appears while you are still recording is a preview of that, and it was not what we scored.

When this was run

5 and 6 August 2026, with AssemblyAI universal, Deepgram nova-3, OpenAI whisper-1, and Aside’s own engine on device. Engines change and so do we. When we measure again, the numbers on this page and on the front page change together, with the date.