A chatbot in Uzbek is not a translated set of buttons but a second language version of the bot: its own texts, its own knowledge base, its own notifications and a separate check of answers on real customer requests. The hard part is not the translation — customers write in Latin script, in Cyrillic and in a mix of Uzbek and Russian, so the language is detected on every message rather than once at the first launch.
Key takeaways
If you sell inside the country, the Uzbek version is needed: most of your customers speak this language, and people write in a chat the way they speak at home. The Russian version stays — business correspondence and part of the requests from Tashkent run on it.
No law states that a bot must speak Uzbek, yet the requirements around the chat already exist.
The sign that a bot without Uzbek already costs you shows up in the dialogues: customers ask in Uzbek while managers answer in Russian or run the text through a translator. The second sign is a person who receives an answer in the wrong language and leaves the chat before placing a request.
How to make Uzbek work not only in the bot but on the website as well is covered in the article on a bilingual site in Uzbek and Russian.
The bot should answer in Latin script and understand both. The direction is set by the state: Cabinet of Ministers Resolution No. 61 of 10 February 2021 approved the roadmap for the full transition to the Latin-based Uzbek alphabet, and since 1 January 2023 record-keeping has been conducted in it.
Correspondence looks different. Some customers type Uzbek words in Cyrillic, others in Latin, and the apostrophe in o‘zbekcha and g‘isht turns into any similar character on a phone: a straight one, a reversed one or a typographic quote. For a knowledge base search these are different words, so the bot reports that nothing was found for a question that does have an answer.
Tip Normalise incoming text to a single form before searching the base: one script, one apostrophe character, double spaces removed. This step alone removes part of the empty answers without a single edit in the knowledge base.
Mixed languages in one message are a separate story: an Uzbek phrase with Russian words inside it, such as dostavka qancha turadi, naloj bormi. There is no need to translate the whole message, a few rules are enough:
The most reliable way is by the text of every message, while the language button and the app language work as hints. The three methods give different results.
| What we compare | Language button | App language | Detection by text |
|---|---|---|---|
| Where the language comes from | The person picks it at the start of the dialogue | From the language_code field that Telegram sends along with the message | From the text of every incoming message |
| What it really is | An explicit choice saved in the profile | The interface language of the app, not the language of the conversation | The actual language of the message together with its script |
| Where it fails | Russian was picked, then the person writes in Uzbek | A Russian interface on someone who writes in Uzbek | Short replies: ha, ok, bare numbers |
| Mixed phrases | Not taken into account | Not taken into account | Answers by the main part of the message |
| Where it fits | Button scenarios: booking and payment | The first greeting before the first question | Free-form questions and support |
A working setup combines all three sources: the app language for the greeting, detection by text on every message, a switch button always in sight. The chosen language is saved in the customer record so that notifications and reminders go out in the language of the dialogue.
Where buttons end and free questions begin is clear from the example of a Telegram bot for client bookings: the booking scenario itself is easy to make bilingual, while questions about what happens if the client is late arrive as free text.
Worse than Russian and English, and the spread between models is much wider. This is visible in UzLiB, an open benchmark of the Uzbek language: questions about correct word forms, meaning and usage were collected by native speakers, and the results table keeps growing with new models.
The last figure matters more than the first: a label saying the model was trained on Uzbek promises nothing on its own, and what has to be checked is the result rather than the description. Proprietary models still answer in Uzbek better than open ones, which the TUMLU benchmark for Turkic languages showed as well: the top five places there went to proprietary models.
Even so, the top rows of any table say nothing about your price list and your terms. A customer asks about delivery to Samarkand and instalments, not about word forms, and the accuracy of the answer is set by the knowledge base and the answering rules. How that is arranged is covered in the article on AI assistants that answer from company documents.
A voice message is a normal way to ask a question in Telegram, and Uzbek speech is transcribed: the language list of Google Cloud Speech-to-Text includes uz-UZ with the Chirp, Chirp 2 and Chirp 3 models, and Russian is supported there as well.
Trouble starts where languages are mixed inside one recording and a street or a sales floor is heard in the background. That is why amounts, dates, addresses and names from a voice message are confirmed by the bot with a question or a button instead of being written into the request silently.
How the share of transcription errors is measured and what to do with bilingual recordings is covered in the article on AI analysis of calls and chats in sales.
The core items are an Uzbek knowledge base, a glossary of terms and a person who continues the dialogue after the bot. Translating button labels is the simplest item on the list, and usually the only one that gets done, after which the bot politely answers in general phrases.
The last item decides the fate of the project. If a handed-over dialogue reaches a manager who does not write in Uzbek, the customer gets the same answer as before the bot, only later.
Testing is done before the bot is opened to customers, and it is counted in numbers rather than judged by eye.
If Uzbek produces noticeably more errors, those topics are handed to a person at the start: the bot answers only where it is reliable and widens its area as the base grows.
The limits are the same as for a Russian-language bot, but the cost of a mistake is higher: there is often nobody inside the company to check an Uzbek answer, and a wrong wording lives in the correspondence for months.
Important Correspondence with a customer is personal data. The Law on Personal Data No. ZRU-547 allows processing on one of the grounds listed in Article 18, most often the person's consent, and requires biometric and genetic data to be stored in Uzbekistan only. If dialogues go to a cloud model abroad, this is a cross-border transfer, and its conditions are checked before the launch.
The Uzbek version does not double the budget: the platform, the scenario and the integrations are the same, what is added is the texts, the second version of the knowledge base and its review. The starting prices on the service page are these:
Three things push the budget above the starting price: the volume of the knowledge base to collect and translate, the number of channels and systems to connect, and work with voice messages. What each format includes is described on the page about bots and AI assistants.
After the launch there remain the payment for model requests or your own server and the work of the person who keeps the knowledge base current in both languages. Support of the company's working systems starts at $700 per month.
Syntra Systems starts with your dialogues rather than with the choice of a model: we look at the questions customers ask in Uzbek, which topics close without a person and what the bot must hand over to a manager. In parallel we collect the knowledge base and the glossary in both languages so that the answers match in meaning and in terminology.
Before opening access we run the bot on real customer requests and count the share of wrong answers separately for each language, and after the launch we read the dialogues and fill the base with the questions that had no answer. Uzbek texts are checked by a native speaker, not by an automatic translator.
Let’s discuss your project
Tell us what you need, and we will estimate the timeline and cost and suggest a solution.
For button labels and short hints that is enough, for prices, terms and instructions it is not: machine translation shifts the wording and does not know your terms. The working option is a knowledge base in Uzbek read by a native speaker, plus a glossary of service and product names.
In Latin: it is the script set as the direction of the Uzbek alphabet transition, and record-keeping is conducted in it. Cyrillic still has to be understood — part of the customers type Uzbek words that way, so incoming text is normalised to a single form before the knowledge base search.
By the share of correct answers counted separately for each language on your own customer requests from the previous period. Mixed Uzbek-Russian messages are counted as a separate group: the bot fails on them most often, and those topics are better left to a person.
Yes. The bot hands over everything on the closed-topic list: discounts, disputed cases, legal and medical questions. If a handed-over dialogue reaches a manager who does not write in Uzbek, the customer gets an answer in the wrong language and loses time.
Uzbek speech is transcribed by cloud recognition services, including the variant for Uzbekistan. Results are worse on noisy recordings and where languages are mixed within one phrase, so amounts, addresses and dates from a voice message are confirmed by the bot.
The platform, the scenario and the integrations are the same, what is added is the texts, the second version of the knowledge base and its review. A bot for requests and bookings starts at $1,200, an assistant answering from company documents at $3,000, and a review of processes with an automation plan costs $400.
Answer in the language of the larger part of the message and avoid resetting the dialogue to a language selection screen. Terms and product names are kept in a glossary with their forms in both languages so that the answer does not mix word forms or distort the name of a service.
Sources
Cover photo: Mahdi Bafande, Pexels
