Why this benchmark exists, and what it is not

There are already Cantonese speech benchmarks, and any claim that this is the first would be false. WSYue-eval covers recognition and synthesis. HKCanto-Eval, CantoNLU, MDCC and Common Voice zh-HK all exist and all predate us. We cross-anchor against WSYue-ASR-eval precisely so that a character error rate published here can be read alongside one published there.

What does not exist is an evaluation on real Hong Kong insurance contact-centre audio: consented recordings of actual calls, with the hesitation, the crosstalk, the background of an open floor, and the code-switching that a Hong Kong caller does without noticing. Read speech is a different problem and models that do well on it do not reliably do well on this.

Four pillars, because a voice system is four systems. Can it hear, can it understand, can it speak, and can it hold a conversation. Entrants may enter any subset. A team that enters two pillars is ranked on the Core Score alongside everyone else, and a pillar not entered is reported as not entered, never as a zero.