Fact · Speech recognition
4.20percent word error rate
In our October 2026 test on a Mac with an M2 Pro, Whisper large-v3-turbo had a word error rate of 4.20 percent on 100 English FLEURS recordings with background talk 20 dB quieter than the speaker (a quiet office).
Measured by us on October 19, 2026, from 100 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.
- 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. The same 100 English recordings as our main benchmark, so the conditions differ only in the noise added.
- How it is measured
- Word error rate = 90 word errors / 2,141 reference words = 4.20 percent. Babble = four other speakers from the same recordings at equal loudness, mixed at 20 dB below the speaker.
- What it does not show
- Babble made from read sentences is steadier than a real cafe (no dishes, music or sudden laughter), and a close microphone in a real room changes the ratio. Compare engines and conditions, not this number with a real room.
In the data
A row of Dictation in Noise: Speech Recognition Accuracy Tested, as Whisper large-v3-turbo, 20 dB: 4.2 percent.
- Engine
- Whisper large-v3-turbo
- Maker
- OpenAI
- Background talk
- 20 dB
- Language
- English