Speed and accuracy, measured

The graphics numbers come from the integrated Radeon graphics of an ordinary laptop, not a gaming card. Even so, the graphics chip ran Standard 5.6 times as fast as the processor, and Accurate 2.5 times as fast. In a separate short test it also heated the laptop less: Standard on graphics raised the temperature reading by 6 °C, the same model on the processor by 15 °C.

With Standard on graphics the first line of text appears within a couple of seconds, so you can start reading while the rest is still coming.

Sustained work ran slower per hour than the short clips did (6.3× against 17×), so plan with the long-file figure for long recordings. Transcribe saves its progress as it goes: a long job survives closing the app or restarting the PC, and continues where it stopped.

Word error rate counts the words a transcript gets wrong (replaced, missing or added) against a reference text, divided by the number of words in the reference. SNR, the signal-to-noise ratio, says how much louder the speech is than the noise: at 20 dB the noise is in the background, at 5 dB it is loud enough to make people strain.

Accurate made fewer mistakes than Standard in every condition, between a third and nearly half fewer, and the gap held as the noise grew. The Russian recording was the hardest file for both models.

Unusual names are where automatic transcription goes wrong most visibly, and where a wrong letter matters most. Given the names in advance as custom words (a Pro feature), the Standard model wrote every one of them correctly. More ways to improve accuracy.

The test set uses synthetic voices, which are easier than real people. We also ran the app's own sample, a public-domain LibriVox reading of the Gettysburg Address (2 min 38 s). Standard got 1.0% of its words wrong, Accurate 7.4%.

One recording is far too small a sample to rank the models, but it is a useful reminder: Accurate is better on average, not on every recording. Try both on your own audio. The free plan includes 15 Accurate minutes for exactly that.

• When and what: November 8, 2026, with the speech engine of Transcribe 1.0: whisper.cpp running Whisper small (Standard) and Whisper large-v3-turbo (Accurate).
• Machine: a laptop with AMD Radeon integrated graphics (Vulkan) and 8 processor threads.
• Test set: six scripted recordings of 41 to 104 seconds (a clinic note, a lecture, a legal intake, a dictation full of names, a research interview, and notes in Russian), made with synthetic voices, each also with background noise mixed in at 20, 10 and 5 dB SNR. Plus a three-hour recording and one human reading.
• Speed: the audio's length divided by the time the engine took, without loading the model. First line: how long until the first line of text appeared.
• Accuracy: word error rate, comparing the transcript word by word with the reference text.
• Memory: the engine's peak working set.
• Every configuration ran once per file.

Download all 76 runs as CSV: model, device, file, condition, audio length, processing time, speed, word error rate with substitutions, deletions and insertions, peak memory and the time to the first line. Use it however you like, and tell us if you measure something different on your PC.

• Start with Standard. With graphics that have a Vulkan driver it keeps far ahead of real time: an hour of audio in a few minutes, free and without limits.
• Use Accurate for noise, accents and names. In our tests it made a third to nearly half fewer mistakes. On the processor alone it runs at about real time, so let long recordings run in the background.
• Add custom words when names matter (Pro). On our names test they removed every error.
• Check what matters against the audio. No model is perfect, and automatic transcription can occasionally write words that were not said. In Transcribe, hearing a line again is one click.