Meet Tsuno at Sirha Lyon 2027 · Stand S.57Learn more
EN FR
Try the demo
Menu

Team and management

Mixed languages in the kitchen, a reality software ignores

Kitchens speak a mix of slang, jargon and home languages. A good tool learns the house words, shows what it understood and never corrects anyone.

Published on

In the middle of prep, two cooks share a laugh over a worktop covered in herbs and containers.

Kitchens do not speak textbook language. They speak a mix of slang, technical terms, house names for dishes and words from several languages, often in the same sentence. Software designed for a standard language fails on this. A useful tool learns the house vocabulary and shows what it understood before acting. It never picks on the way someone speaks.

Saturday, 8:05 pm, at the pass

"Send the 12, ya khouya, we're out of chives!"

Rachid has been on the hot line since 6 pm. He has just said three things in one breath: table 12 can go, he is talking to Corentin, who runs garde manger, like a brother ("ya khouya" is colloquial Arabic), and the cold station is out of an herb. Corentin got all of it in a second. Thanh, the commis he is training, caught "chives" and is already reaching for the tub.

Nobody in this kitchen finds the sentence strange. It is the language of service: fast, warm, full of shortcuts. It works because everyone shares the context.

Now picture the same sentence spoken to a consumer voice assistant. It looks for a song, or says it did not understand. (Characters and venue are fictional.)

Why software fails on kitchen language

It expects standard language

Most tools are designed and tested on "clean" language: full sentences, everyday vocabulary, one language at a time. Kitchens work the other way round. Clipped sentences, missing verbs, two-word orders. "Two steaks, one rare" is perfectly clear at the pass and almost meaningless out of context.

It does not know the house vocabulary

Every venue has its own words. The signature dish has a nickname. "The back tub" means one particular tub. A prep is named after the chef who invented it. A tool that does not know these words swaps them for the nearest word in its dictionary, which is usually one that does not exist in the room.

It treats mixing as an error

Switching languages mid-sentence has a name among linguists: code-switching. In multilingual teams it is the normal way to talk. An Arabic word for warmth, a Portuguese term nobody ever translated, an English "check" in a French kitchen. A tool built for one language at a time splits these sentences badly, or rejects them.

It understands some accents less well

It has been measured. A Stanford study published in PNAS in 2020 found nearly twice as many transcription errors for Black speakers as for white speakers across five commercial systems; we cover it in accents, stuttering, hearing loss. It concerns American English, yet the lesson carries over: a voice tool's errors land first on people who speak differently from its training data.

What a tool needs to understand

  1. The house vocabulary. Dish names, preps, stations and equipment, as the team actually says them. A good tool relies on the venue's own recipe sheets and procedures, not on a generic dictionary.
  2. Short, elliptical sentences. "The 12, good to go?" has to be understood as a question about a table, not as a riddle.
  3. Mixed-language sentences. A request that mixes two languages should be handled as a request, not as a mistake. This is the hardest part, and the one to test most.
  4. Variants of the same word. "Chives", a shortened slang version, the word in another language: if the team says all three, the tool must cope.
  5. Its own uncertainty. When it is not sure, the tool says so, shows what it understood and lets the person correct it. Being able to say "I don't know" beats an invented answer.

What a tool must never do

  1. Correct grammar. A cook who says "we're out of chive" does not need to hear "chives". They need chives.
  2. Judge a way of speaking. No diction score, no fluency rating, no comparison between employees. Understanding a request is not assessing the person who made it.
  3. Mock an accent or ask for repetition on a loop. "Sorry, I didn't catch that" three times in front of the team is humiliating. After the second miss, the tool offers the screen.
  4. Infer anything from the voice. Not mood, not origin, not tiredness. In the European Union, inferring emotions in the workplace has been banned since 2 February 2025 (Regulation (EU) 2024/1689, article 5), except for medical or safety reasons.
  5. Act on a misheard sentence. Any action has to be summarised and confirmed: if the tool heard "table 2" instead of "table 12", the person must see it before anything goes out.

What a tool gets right, what it must avoid

SituationPoorly designed toolGood tool
"Send the 12, we're out of chives""I didn't understand your request"Recognises the table and the ingredient, asks for confirmation if an action is involved
House nickname for a dishSearches for an unknown dish or offers a generic recipeFinds the venue's recipe sheet under that name
Sentence half in English, half in ArabicHandles only the English part, loses the meaningHandles the whole request, or says what it did not understand
Strong accent, extractor noiseAsks for repetition three timesShows what it understood, offers to correct on screen
"The 12, good to go?"Unrelated answerUnderstood as a question about table 12
Grammar slipRephrases or correctsAnswers the question, without comment
Allergen questionAnswers from memoryShows the venue's approved sheet, or refers to a manager

Common mistakes

  • Testing a voice tool in the office, in quiet, in formal language. A kitchen on Saturday night is another language.
  • Asking the team to "speak properly" to the tool. The tool should adapt to the team, as argued in technology should adapt to people who work.
  • Forgetting the house words. If the sheets use one name and the team says another, the tool will be lost. Record both.
  • Taking a slick demo as proof. Only your own trials, with your voices, your noise and your vocabulary, show whether it works for you.
  • Treating language as a fault. A mixed sentence often carries a team's closeness, and correcting it breaks that.

What to measure

  • During a trial, the share of requests understood first time, recorded by station and by moment of service, never by person.
  • The number of abandoned requests: the person gave up, or asked a colleague instead.
  • A list of house words the tool got wrong, updated as trials go on. It should get shorter.

Where Tsuno San comes in

Tsuno San is the intelligent voice assistant for teams, connected to the Tsuno environment: procedures, tasks, checklists, checks, equipment and venue information. That connection is what gives it access to the house words: the names of your sheets, your stations and your equipment, as you wrote them.

We are working so that San can converse in French, English, Arabic and Vietnamese. Sentences that mix several languages are something we are working on: to be tested, not a given. We do not publish an understanding rate; we would rather you see it for yourself in a demo, with your team's own words.

What is already true: San Voice, the voice interaction, is available in demo. San acts within the person's permissions and launches nothing without their confirmation, as what San never does explains. It assesses nobody and does not try to guess anyone's mood or motivation. The data rules are on the Trust and data page.

Further reading

For the manager doing the welcoming, read onboarding a multilingual team: a manager's guide. On the voices tools understand less well, see accents, stuttering, hearing loss: a voice tool for everyone, or for no one. To run a serious trial, how to test a voice assistant in a kitchen, and to see why a general-purpose tool gets lost in your venue, why a general AI does not know your restaurant. For an overview of the tool, read the pillar page Tsuno San, the work companion for frontline teams.

Key takeaways

Kitchen language is made of slang, jargon and several home languages, and it works. A good tool learns it and shows what it understood, without ever correcting anyone.

Frequently asked questions

Why do voice assistants struggle with kitchen language?

Because they are built for a standard language. Kitchens mix slang, technical terms, house names for dishes and words from several languages in the same sentence. Without that vocabulary, a tool guesses, and gets the important words wrong.

What is code-switching?

It is moving from one language to another in the middle of a conversation or a sentence, for example from French to Arabic or from English to Spanish. Linguists call it code-switching. In multilingual teams it is a normal way of speaking, not a mistake.

Should a work tool correct the way people speak?

No. A work tool should understand the request, show what it understood and let the person correct it. Fixing grammar, commenting on an accent or asking people to repeat themselves again and again humiliates them and puts them off using it.

Are accents a problem for speech recognition?

They can be. A study published in PNAS in 2020 measured, across five commercial systems, a word error rate almost twice as high for Black speakers as for white speakers. A tool should therefore be tested with the real voices of your team.

Does Tsuno San understand sentences that mix languages?

That is something to test, not a given. We are working so that San can converse in French, English, Arabic and Vietnamese, and mixed-language sentences are part of our trials. The best way to judge is a demo using your own team's words.