Preprint
The Edinburgh International Accents of English Corpus: Towards the Democratization of English ASR
Ramon Sanabria et al., arXiv (accepted to IEEE ICASSP 2023), 2023
Read the originalRead on October 17, 2026
- What it is
- A corpus of almost 40 hours of video calls between friends, recorded across many varieties of English, which the authors used to test public and commercial English speech recognition models. It shows how far average accuracy falls for several accents compared with US English read speech.
- How much weight it carries
- Peer-reviewed conference paper, released with its recordings and evaluation scripts under CC-BY-SA. Conversational speech is harder than read speech, so the comparison with a US clean-read baseline is not like for like. The per-accent results are shown in a figure, so this library uses only the averages and the named accent groups from the text.
Facts taken from this source
- On the EdAcc accent test set of conversations between friends, the best-performing English speech recognition model, trained on 680 thousand hours of transcribed data, averaged a word error rate of 19.7 percent, against 2.7 percent on US English clean read speech.
“The best performing model, trained on 680 thousand hours of transcribed data, obtains an average of 19.7% word error rate (WER) -- in contrast to the 2.7% WER obtained when evaluated on US English clean read speech.” (Abstract)
- In the EdAcc study, every English speech recognition model tested showed a drop in performance on speakers of Indian, Jamaican and Nigerian English.
“Across all models, we observe a drop in performance on Indian, Jamaican, and Nigerian English speakers.” (Abstract)
- The same best English speech recognition model, tested by the EdAcc authors on US English clean read speech, had a word error rate of 2.7 percent.
“obtains an average of 19.7% word error rate (WER) -- in contrast to the 2.7% WER obtained when evaluated on US English clean read speech.” (Abstract)