Fact · Speech recognition
13.50percent 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 13.50 percent on 83 conversational utterances by Chinese speakers (2 speakers) from the EdAcc corpus.
Measured by us on October 21, 2026, from 83 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 recordings are English conversations over video calls between people with different accents (EdAcc, University of Edinburgh, 2023), not read sentences, so error rates run higher than on our read-speech benchmark. The utterances were picked by a fixed rule before any engine ran: at least five words, 1.5 to 30 seconds, no overlapping speech, up to 100 per accent group spread across its speakers.
- How it is measured
- Word error rate = 309 word errors / 2,289 reference words = 13.50 percent, over 83 utterances; hesitation fillers are left out of both texts and ok counts as okay, for every engine alike.
- What it does not show
- 2 speakers stand for this group, so one speaker's voice, microphone and topic weigh heavily: compare engines within a group more than groups with each other. Numbers spoken as words in the reference count as errors when an engine writes digits.
In the data
A row of Speech Recognition and Accents: 13 English Accents Tested, as Whisper large-v3-turbo, Chinese speakers: 13.5 percent.
- Engine
- Whisper large-v3-turbo
- Maker
- OpenAI
- Accent group
- Chinese speakers
- Kind of accent
- First language other than English
- Speakers
- 2
- Utterances
- 83