In our October 2026 test on a Mac with an M2 Pro, Whisper large-v3-turbo had a word error rate of 2.91 percent on 84 English readings of 3 to 4 minutes by 40 speakers, handed over as one recording of 3 to 4 minutes.

Measured by us on October 25, 2026, from recordings

Our own measurement

This number comes from our benchmark, not from another publication. The full record (the setup, every recording's checksum, every transcript and timing) is published, so you can recompute it.

Dictation Accuracy on Long Recordings, Tested Download the full record (JSON)

What it measures: OpenAI Whisper large-v3-turbo, 5-bit quantized (ggml-large-v3-turbo-q5_0.bin), run through whisper.cpp 1.9.4 with Metal on the GPU, with the language given. Whisper ran with greedy decoding and timestamps on, which is how whisper.cpp moves through audio longer than its 30-second window; whisper.cpp's beam search was not tested. 84 examples of 3 to 4 minutes read by 40 speakers from public-domain audiobooks (LibriSpeech-Long test-clean, Google DeepMind, from LibriSpeech), clean recordings of read English. In the pieces condition the same audio was cut at pauses into pieces of 10 to 20 seconds, as someone dictating in short bursts would, and the pieces' text was joined.
How it is measured: Word error rate = 1,549 word errors / 53,292 reference words = 2.91 percent, over 5.45 hours of audio. 95% interval over speakers: 2.54 to 3.27 percent. Left-out words: 0.55 percent of the reference; examples that lost a passage of 10 or more words in a row: 4 of 84.
What it does not show: The same audio in pieces gave 2.35 percent; the difference (one recording minus pieces) is +0.56 points, with a 95% interval allowing for the test's 3 such comparisons of +0.23 to +0.91 points, which excludes zero, so it is a measured difference. Audiobook reading in a quiet room is easier than dictating your own words at a desk, and an app may cut or stream audio in its own way; the test isolates what the length of one recording costs each engine.

In the data

A row of Dictation Accuracy on Long Recordings, Tested, as Whisper large-v3-turbo, in one recording: 2.54 to 3.27 percent.

Engine: Whisper large-v3-turbo
Maker: OpenAI
How the speech was handed over: One recording of about 4 minutes
Readings: 84
Speakers: 40
Language: English

Used to calculate

The speech recognition accuracy on long dictation: one 4-minute recording against the same speech in short pieces dataset holds 6 data points, each a published figure traced to its source.