Voice POS

A waiter holds the mic and says “Table twelve, two colas, one sea bass, a 20 cl rakı.” Speech is transcribed by Whisper running on the Apple GPU, a deterministic Turkish parser turns it into a table and line items, and an approval screen shows the interpreted order before it is added.

The Problem

In a busy restaurant, tapping through menus on a POS is slow and error-prone. Voice is faster — but a POS cannot guess: a wrong table or a misheard dish costs real money.

The Solution

Fast local speech recognition plus a parser that never improvises: exact, alias and fuzzy matching against the real menu, then a human confirms. Recognition runs on the host GPU because Docker on macOS cannot accelerate it — about 0.4 s instead of 11 s.

What I Built

• →Whisper large-v3-turbo on MLX, kept alive as a macOS service
• →Table auto-detected from speech, or picked by tap
• →Turkish order parser with exact, alias and fuzzy matching — no AI tokens spent
• →Approval modal to review and edit before the order lands
• →Roles for admin, waiter and kitchen display
• →Demo transcript button so the whole flow works without a microphone

Highlights

• →~0.4 s per utterance on the Apple GPU
• →No LLM in the parser — zero token cost
• →Approve-before-add safety step

Outcome

• →Speech-to-order in well under a second per sentence on local hardware
• →Live demo at smartpos.mehditerzi.com