Generic AI knows the language. ShopVoices knows your shelves.
A shopper doesn't ask for "a cordless vacuum cleaner, model VX11-2" — they ask for "sådan en ledningsfri støvsuger". The Console tunes speech recognition and understanding to your catalogue — your brands, your product names, the way your shoppers actually talk — and proves every improvement with real test suites before it reaches the floor.
Eight guided stages from catalogue to verified quality
Recognition readiness is a checklist, not a consultancy engagement. The wizard walks through eight stages — catalogue in, vocabulary generated, languages configured, aliases reviewed, test set built, quality measured — and shows exactly which stages are green and what the next step is. Anyone on the team can read where a store stands.
- Each stage explains itself and links straight to the work.
- Readiness is visible per store on the Console's front page.
- An automated optimizer can run the routine stages on its own.
Your product names become part of the recognizer
The Console reads your catalogue and generates the recognition vocabulary and corrections from it — the brand names, the model lines, the Danish compound words a generic system mishears. You review what it generated, then publish; every device in the store picks it up. When the catalogue changes, regenerate. The assistant literally speaks your assortment.
- Vocabulary and corrections generated from your own product data.
- Review before publish — nothing changes on the floor unseen.
- Danish-first: built for the compounds and accents generic systems miss.
Teach it that "hoover" means vacuum cleaner
Every chain has its own shopper vocabulary — nicknames, abbreviations, the competitor's brand used as a generic word. The Aliases tab holds them all, with AI-suggested candidates you approve or reject, and the Phrases tab curates the follow-up grammar: which "anything cheaper?"-style refinements the assistant offers and understands.
Every change is benchmarked before it ships to the floor
The test suite holds labelled recordings — real questions with known right answers — and runs them through the actual recognition pipeline. Change the vocabulary, run the suite, compare against the baseline: better ships, worse doesn't. Recognition quality stops being an opinion.
- Labelled test sets per store and per language.
- Baseline-vs-candidate comparison as the no-regression gate.
- AI-generated test samples to grow coverage quickly.
The understanding report closes the loop
Which languages are actually spoken at each kiosk? Where does recognition hesitate? How fast is the pipeline, stage by stage? The understanding report shows confidence, language mix, response times and the questions that struggled — so tuning effort goes where the floor needs it, not where someone guessed.
Hear it speak your assortment
A pilot includes tuning to your catalogue — and the test suite that proves it worked.