Skip to content

Fact · Speech recognition

5.18percent 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 5.18 percent on 100 English FLEURS recordings after the narrowband voice path of the Bluetooth Hands-Free Profile (CVSD codec, 8 kHz), the one every hands-free headset must support.

Measured by us on October 22, 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.

Bluetooth Headset Dictation Accuracy, TestedDownload 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. The same 100 English recordings in every condition, so the conditions differ only in the audio path.
How it is measured
Word error rate = 111 word errors / 2,141 reference words = 5.18 percent. Every recording was brought to the same speech level (-26 dB below full scale) and, for this condition, encoded and decoded in software exactly as described in the run record.
What it does not show
Compared with the original recording, this engine's word error rate changed by +1.26 points. Over the 100 recordings, the 95% interval for that change, allowing for the test's nine comparisons, is +0.28 to +2.93 points, which excludes zero, so this is a measured change. The codecs are run in software on studio-clean read speech. A real headset adds its own microphone, noise suppression and radio losses, which this test leaves out, so compare conditions and engines rather than this number with a headset in a room.

In the data

A row of Bluetooth Headset Dictation Accuracy, Tested, as Whisper large-v3-turbo, Narrowband (CVSD): 5.18 percent.

Engine
Whisper large-v3-turbo
Maker
OpenAI
Audio path
Narrowband (CVSD)
Language
English

Used to calculate

Cited on