Fact · Languages and accents
16.5percent character error rate
On the third evaluation set of the Corpus of Spontaneous Japanese (CSJ eval3), Whisper large-v3 had a character error rate of 16.5 percent in the paper's comparison table.
Checked against its source on October 17, 2026
In the source's own words
“Table 3: Comparison on CSJ data (CER %). size eval1 eval2 eval3 Whisper large-v3 1.55B 18.4 17.5 16.5”
Whale: Large-Scale multilingual ASR model with w2v-BERT and E-Branchformer with large speech data, Yosuke Kashiwagi et al., arXiv (Interspeech 2025), 2025, Table 3, comparison on CSJ data.
Open the original- What it measures
- The CSJ contains recordings of spontaneous Japanese speech, with fillers and hesitations. The paper reports character error rate for each of three evaluation sets, and the eval3 column is the one given here.
- What it does not show
- Spontaneous speech with fillers and hesitations is harder than typical dictation, which is usually short and deliberate. The table is from the authors' own comparison on their chosen test sets.