Voice ordering in four languages: what we learned
Shopkeepers do not want to type. Building an ordering system around speech in Arabic, English, Hindi and Urdu taught us more about retail than about models.
Typing an order on a phone, in a second language, while serving customers, is not a realistic ask. So we built ordering around the thing every shopkeeper already does effortlessly: speaking.
The hard part was never transcription. It was matching what someone actually says to what is in the catalogue.
People say "two dozen cola" not "Cola 330ml × 24". They mix languages inside one sentence. They say "tissue box" and mean a specific brand.
What worked was less about the model and more about the matching: handling plurals, mapping synonyms, ignoring packaging words, and scoring candidates by how much of the phrase they actually cover. The result is a shopkeeper listing ten items in one breath and seeing ten correct lines appear.