Showcase · Cyberport, Hong Kong
Every entrant’s result is already public the moment it is scored. The showcase is not a reveal, it is the argument: what the numbers mean, where they disagree, and what Cantonese voice can and cannot do yet.
The order is fixed; the times are not, and are Cyberport’s to announce. Rather than print placeholder ones that look real, this page carries the shape and will carry the schedule when there is one.
Standings as of a fixed moment, so the number on the screen does not change while someone is talking about it. The live board resumes afterwards.
The findings, each graded by how strongly the evidence backs it. Where entrants are tied, why they are tied, shown rather than asserted.
Each team presents their system in their own words. We publish the measurements; how they got there is theirs to explain.
A live argument about what these numbers do not capture — the failure modes a corpus of 79 clips cannot reach, and what the next edition should measure instead.
Published the same day, with the working. Nothing said on stage will be absent from it.
Not slides of logos. Real calls, real transcripts, and the specific places where good systems still fail.
The same syllable as six words, played, and a system getting one of them wrong on a real claims call.
Time to first byte on the same conversation across entrants — the difference between a call you stay on and one you hang up.
What each system does when a caller interrupts. The fast ones are not always the ones that stop.
Numbers we refused to publish, and why. A benchmark is worth reading because of what it withholds.
Places are limited and the room is the point — this is a working session, not a launch. Tell us who you are and we will send the schedule when it is fixed.
Cantonese Voice Benchmark 2026.1 · Organised by Cyberport with Prudential Hong Kong. Platform by Pubrio.