Speech recognition, tuned for how Australia talks
The first job is turning sound into words, and it's where generic systems fall down.
A model trained mostly on American English hears "Woolloongabba" or "Prahran" as a
string of bad guesses. Bella's recogniser is tuned for Australian English and the
suburb and street names guests actually say — so the transcript that everything else
depends on is right from the start.
Understanding, not keyword-matching
"Have you got anything for four-ish around eight?" isn't a keyword — it's a request
with a party size and an approximate time. Bella reads it as one intent, extracts the
details, and fills in the blanks by asking, the way a good host would. There's no menu
to navigate and nothing to press.
Checked against your real diary
Before Bella offers a time, she reads your live availability — so she only ever
proposes a table that genuinely exists, and the confirmed booking writes straight to
your dashboard. No double-bookings, no reconciling scribbled notes at the pass.
A human in the loop, by design
This is the part that makes an AI safe on a real phone line. Every response carries a
confidence score. When it's high, Bella handles the whole call. When it drops — a
tricky request, a name she can't quite catch — she stops guessing, takes the caller's
details, and flags your team to ring back. A low-confidence moment becomes a callback,
not a wrong booking.
Fast enough to feel human
None of this is useful if it's slow. Bella answers in under a second and keeps the
conversation moving at a natural pace, so callers get the brisk, friendly experience
they'd want from your best host on a quiet night — not a laggy robot.
Processed onshore, and kept in Australia
Call audio is processed in Australia, retained only as long as it takes to confirm the
booking, and never used to train outside models. Biteperk is Privacy Act compliant, and
the bookings — and the recordings — stay yours.