Cantonese Voice Benchmark粵語語音基準

E2E · 25% of the score · optional

Can you converse?

Audio in, audio out, on the clock. Not whether the parts work, but whether the whole thing holds a conversation someone would stay on.

The gap is the product.

Everyone agrees that under 300ms feels instant, and nobody knows what that is like. Play a row: the caller finishes, the wait happens at full speed, then the system starts. Same sentence every time — only the silence changes.

Scored on the 95th percentile, not the median: a system that is usually quick and occasionally takes 2 seconds is experienced as a slow one.

What is measured

Time to first byte

From the last frame of the caller’s audio leaving our socket to the first byte of the reply arriving. Our monotonic clock only — participant timestamps are never trusted. Reported with the measured network round trip alongside, so hosting location is visible rather than silently rewarded.

Conversational naturalness

Human evaluators rate a whole exchange. Did the turn-taking work. Would you trust this with a claim.

Yielding the floor

Every scenario ends on a turn where the caller talks over the reply. We measure how long the system keeps speaking after that, and count it as yielding if it stops within 500ms. Nothing in the protocol asks it to stop — a request would measure compliance with our wire format rather than whether it behaves like a person on a phone.

Why it is hard here

The silence
The gap after you stop speaking is what makes a voice bot feel broken, and it is the first thing a caller notices.
Turn-taking
Knowing when the caller has finished, in a language with different prosodic cues from English.
Compounding errors
A mishearing becomes a wrong answer becomes a wrong spoken sentence. The pillars multiply rather than add.
Being interrupted
A system that is fast and will not stop talking is worse than a slow one. Latency-only benchmarks score it as excellent, which is why this pillar tests it explicitly.

A sixty-second clip allows 30000ms from end of utterance to first response byte. That deadline is also the integrity control: it is far too short for a person to transcribe the audio by hand.