auto ts = parakeet::tdt_greedy_decode_with_timestamps(model, encoder_out, cfg.durations);
This does not mean confusables.txt is wrong. It means confusables.txt is a visual-similarity claim that has never been empirically validated at scale. Many entries map characters to the same abstract target under NFKC decomposition (mathematical bold A to A, for instance), and the mapping is semantically correct even if the glyphs look nothing alike. But if you treat every confusables.txt entry as equally dangerous for UI security, you are generating massive false positive rates for 96.5% of the dataset.
。关于这个话题，服务器推荐提供了深入分析
I also put "-" between things that look like multiple words, but would be written as one word today.
無料で日本語もサポートしリアルタイム音声アプリをWhisperより高精度で開発できるオープンソースAIツールキット「Moonshine Voice」