<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Aiphoria Team on Medium]]></title>
        <description><![CDATA[Stories by Aiphoria Team on Medium]]></description>
        <link>https://medium.com/@aiphoriateam?source=rss-7c54839448b6------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*T_4Q8VGqDUT5ROHcIUFUbw.png</url>
            <title>Stories by Aiphoria Team on Medium</title>
            <link>https://medium.com/@aiphoriateam?source=rss-7c54839448b6------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 16:52:04 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@aiphoriateam/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[When Languages Lack Data: Making Open-Source LLMs Work for Banking in Uzbekistan]]></title>
            <link>https://medium.com/aiphoria/when-languages-lack-data-making-open-source-llms-work-for-banking-in-uzbekistan-663e6b3891b8?source=rss-7c54839448b6------2</link>
            <guid isPermaLink="false">https://medium.com/p/663e6b3891b8</guid>
            <category><![CDATA[banking-technology]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[low-resource-language]]></category>
            <category><![CDATA[uzbekistan]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Aiphoria Team]]></dc:creator>
            <pubDate>Mon, 02 Jun 2025 11:40:10 GMT</pubDate>
            <atom:updated>2025-06-02T11:44:14.275Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*K4Hq7xCPtxQ7bn5Z.jpeg" /></figure><p>At Aiphoria, we develop AI agents designed to automate employee work across diverse organizational environments. Our <a href="https://aiphoria.ai/">Aiphoria Pro Platform</a> is a next-gen NLP solution that easily integrates AI into any business process — text or voice-based. It adapts to any knowledge domain or language with open architecture and on-premise deployment options. Our GOAL framework (Goal-Oriented Architecture for LLM) ensures an AI employee guides customers toward specific objectives rather than through rigid rule-based scenarios, whether selling products or solving technical issues.</p><p>TBC Uzbekistan, <a href="https://techcrunch.com/2024/12/20/uzbekistans-mobile-bank-tbc-bags-37m-to-expand-with-new-ai-and-insurance-products/">the dominant financial player</a> not only in Uzbekistan but across Central Asia, embarked on an ambitious vision: developing proprietary Speech Technology in both Uzbek and Russian languages — arguably the most ambitious AI project in the entire Central Asian region. To realize this vision, TBC Uzbekistan assembled a powerful internal team with specialized expertise and strategically partnered with our company, whose experience and technological capabilities would help bring their ambitious plans to life.</p><p>The joint project underscored a critical challenge in AI integration: the limitations of standardized, existing AI solutions in complex, diverse markets. Our work began with developing advanced Automatic Speech Recognition (ASR) and Text-to-Speech (TTS) technologies for the low-resource Uzbek language, which we’ve covered extensively <a href="https://medium.com/aiphoria/ai-agents-in-uzbekistan-cad2f5b3c1b3">in our previous post</a>. However, it’s not enough to just speak or hear Uzbek — you need to teach your system to understand what people are saying and respond appropriately.</p><p>Thankfully, Large Language Models (LLMs) can help, but the challenge persists — Uzbek is still a low-resource language, meaning there are no ready-to-use solutions. This is especially true in banking, our target domain. Banking requires LLMs not just to speak Uzbek but to understand specialized vocabulary. Customers won’t discuss weather or news with a banking AI — they’ll ask about loans, debts, and credits in all their complexity. Banking also requires any LLM solution to work on-premise due to high security and data privacy concerns. Additionally, banking interactions demand speed. Most communications happen by phone, with customers frequently interrupting, mishearing, mixing languages and dialects. For our LLM solution to work effectively in voice applications, it needs to be exceptionally quick.</p><p>Once again, we face an interesting technological challenge: creating a tailored AI solution for a culturally nuanced market — this time, leveraging LLMs in true partnership with TBC Uzbekistan’s expertise. The AI team at TBC Uzbekistan collaborated closely with us, sharing their insights on local banking practices and customer communication patterns that significantly enhanced our solution’s effectiveness. Their technical team was also instrumental in the implementation phase, ensuring seamless integration with existing banking systems and security protocols — a critical aspect that can make or break AI adoption in financial institutions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*SqCvTssqcIzcTZ8r" /><figcaption>Image credit: TBC Uzbekistan</figcaption></figure><p>So let’s dive deep into how we’ve managed this all.</p><h3>What we had to keep in mind</h3><p>One obvious option would be training a model from scratch. However, it would require massive datasets: billions of texts in all required languages, including Wikipedia. On top of that, the computational resources needed for training would be significant to say the least. For context, training a model like GPT-3 is <a href="https://www.visualcapitalist.com/training-costs-of-ai-models-over-time/">believed to cost</a> millions of dollars. It wouldn’t make much commercial sense to spend such an amount of money on this task. Thankfully, there is always a smarter way.</p><p>So, instead of starting from scratch, we decided to fine-tune existing open-source LLMs. As we’ve already explained in our previous post, Uzbekistan is a country where several languages and dialects are spoken. This means that we needed a <strong>fully bilingual stack </strong>— LLMs that could understand and generate text in both <strong>Uzbek </strong>and<strong> Russian</strong>. Additionally, we wanted to deploy the system <strong>on-premise </strong>because of all the data safety concerns as we’re talking about banking. Our target was models that would also be:</p><p>– small enough to run well on our client’s appropriately cost efficient hardware without the need to upgrade it by pouring additional millions into infrastructure</p><p>– fast enough to handle quick responses when working in voice modality during phone calls</p><p>– powerful enough to handle the complexities of bilingual text generation</p><p>All of this makes our target model about <strong>7–9 billion parameters </strong>large. Although there is no universal definition of what is a small model and what is a large model, small models are usually below 1 billion parameters while very large models like GPT-4 can reach more than 1 trillion parameters. This makes our target model relatively small.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*n5QquXlzCSQC1ok_G8Hx0g.png" /><figcaption>An example of how LLMs can be categorized</figcaption></figure><h3>Choosing the Right Model</h3><p>Nearly all open-source models are released in two versions: pretrained (-pt) and instruction-tuned (-it). Pretrained models have a single capability — predicting the next word based on previous context. They understand that after “the cat sits on the” comes “mat” rather than “mats,” or that after “heavy” comes “rain” not “raining.” Essentially, pt-models learn language structure.</p><p>However, most users need models to follow specific instructions: “answer a question,” “tell a joke,” or “choose from the given options.” Pretrained models can’t follow these instructions, so they undergo additional training on instruction datasets to create instruction-tuned (-it) models. In other words, instruction-tuned models evolve from pretrained ones.</p><p>Our goal was to create an instruction-following assistant capable of responding in two languages. This presented several strategic choices: should we start with a -pt or -it model? Should we use multiphase training (completion and instruction + alignment) or single-phase (instruction only)? Should we implement full fine-tuning (retraining the entire model or specific layers) or parameter-efficient fine-tuning (like LoRA)? Should we separately train the embedding layer or leave it unchanged?</p><p>Different combinations of answers to these questions lead to dramatically different training strategies. For example, starting with a -pt model would require extensive training on our target languages followed by comprehensive instruction training (since it hasn’t previously learned to follow instructions). This approach might deliver superior language understanding, but would demand thousands to tens of thousands of high-quality “instruction-response” pairs. These instructions would need to be diverse, including not only domain-specific banking instructions but also, for instance, coding tasks.</p><p>Safety considerations also factored into our decision-making. Instruction-tuned models are typically already trained to safely respond to provocations and inappropriate content requests. Training instructions from scratch with a -pt model would require teaching safety protocols from square one — demanding even more resources.</p><p>Alternatively, using an -it model presents different challenges. Training it on completion tasks might cause it to forget some of its instruction-following capabilities. Skipping completion training and proceeding directly to instruction fine-tuning in our target languages risks the model never fully learns language structure.</p><p>As for choosing between full fine-tuning and parameter-efficient one (PEFT), full fine-tuning specific layers or the entire model requires more time and data. While the exact figures depend on model size and the number of parameters being retrained, the differences can be substantial. Using PEFT methods like LoRA, however, can limit the model’s ability to adapt to new languages, especially if the base model has limited exposure to the target language or related languages from the same language family.</p><p>Each decision represented a trade-off between performance, resource utilization, and development timeline. For a banking application, we couldn’t afford to compromise on accuracy, but we also needed to be practical about computational requirements and development time. The key question became: how could we achieve the best language capabilities and instruction-following behavior with reasonable computational resources and within a practical timeframe?</p><h3>Experimental Approach</h3><p>Our journey began with extensive testing of several model families including Meta’s LLaMa-3.x, Google’s Gemma-2, and Alibaba’s Qwen. We evaluated models ranging from 8B to 72B parameters across four languages. Uzbek and Russian were our target languages, while English established our performance benchmarks. We also included Turkish as a high-resource relative of Uzbek to provide context for Turkic language performance.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/743/0*01IXTyDGqp1bKUBL" /><figcaption>Relative performance of models on Turkish-translated datasets.<strong> </strong>The metric used is a weighted sum of metrics across each model. Normalization is performed relative to the best-performing model.</figcaption></figure><p>Among models under 10B parameters, Gemma-2–9B demonstrated superior performance, particularly outperforming comparable models like LLaMa-3–8B and LLaMa-3.1–8B in Turkish and Uzbek. This clear advantage made Gemma-2–9B our foundation for further training.</p><p>The path to optimal performance wasn’t straightforward. We conducted multiple experiments to determine the best training strategy for our specific needs. We tested vocabulary handling with and without expansion and compared full fine-tuning against more efficient LoRA and QLoRA adapters. Our experiments also explored different training phases, contrasting two-phase approaches (completion followed by instruction) against single-phase instruction-only training.</p><p>Additionally, we investigated the impact of instruction diversity by comparing broad instruction sets against domain-specific banking instructions. As we’ve explained before, base model selection was another critical factor, so we evaluated pretrained versions against instruction-tuned variants to determine our starting point.</p><p>After extensive testing, we developed a surprisingly effective approach. We started with Gemma-2–9B-IT (the instruction-tuned version) and implemented a two-phase training process. The first phase focused on completion training using mixed Uzbek and Russian content, establishing language fluency. The second phase introduced instruction fine-tuning specifically for domain-specific banking scenarios.</p><p>Throughout both phases, we employed Parameter-Efficient Fine-Tuning (PEFT) with QLoRA, which significantly reduced computational requirements while maintaining performance.</p><h3>Completion Training: Learning a Low-Resource Language</h3><p>For the completion phase, gathering appropriate data presented unique challenges for each language. Russian, being a high-resource language, offered abundant datasets. Uzbek, however, required more creative approaches. Models trained on purely translated texts tend to adopt the stilted patterns of translation tools rather than natural language flow, so that wasn’t an option.</p><p>We combined multiple data sources to create a comprehensive training corpus: publicly available Uzbek datasets from Hugging Face, open English and Russian datasets that we translated to Uzbek, and proprietary data from our banking partner, which included anonymized real conversations from support scenarios. These client-provided datasets were essential to making future interactions sound as humane and professional as possible — real data on pre-train helps to learn domain-specific vocabulary and conversational style. This also saved us a lot of time — without them it would’ve taken much more work to learn to answer smoothly.</p><h3>Instruction Fine-Tuning: Teaching Command Following</h3><p>The second phase presented even greater challenges. Despite Gemma being top-tier, neither it nor other instructive models of the necessary caliber can sufficiently handle domain-specific instructions in Uzbek. Which means we needed to create comprehensive instruction-response datasets in two languages — both Uzbek and Russian. Our instruction datasets combined two approaches:</p><ol><li>General instructions (like “generate a question based on a statement”, “write a short story”, “rewrite this sentence to make it more formal”), which we primarily translated from existing resources</li><li>Domain-specific banking instructions, which we both synthetically generated using larger open models and created with human banking experts from TBC Uzbekistan; we also adapted these instructions according to future popular use cases of our solution, such as support scenarios, which means <a href="https://www.pinecone.io/learn/retrieval-augmented-generation/">RAG</a> (Retrieval-Augmented Generation) would be used and instructions should be optimized for it</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/570/0*KvAASwDp_d4l7hgi" /><figcaption>Distribution of automatic translation quality scores for datasets translated from English to Uzbek</figcaption></figure><p>The metric used is a weighted sum of metrics across each model. Normalization is performed relative to the best-performing model.</p><p>Translating instructions proved unexpectedly difficult. Many standard LLM training instructions contain language-specific elements that might lose meaning when translated. For example, instructive tasks such as disambiguation resolution lose their sense after translation because homonymy disappears due to translation.</p><p>In order to avoid it, rather than relying solely on automated translation we engaged native Uzbek speakers to review our translated instructions. This human oversight revealed that approximately 70% of translated instructions maintained their original meaning in Uzbek and grammar and syntax were also correct, which proved that the underlying idea of automated translation was viable. For the problematic 30%, our language experts helped reformulate instructions to preserve their training value while making them culturally and linguistically appropriate wherever possible. However, not everything could be fixed without losing the correct meaning, so some instructions had to be ditched, which was by itself also a valuable contribution from human experts.</p><p>Now to synthetic data of domain-specific instructions, another important part of our fine-tuning process, which also came not without some caveats. Models trained primarily on synthetic data perform well with similar synthetic inputs but struggle with authentic human interactions. Early live testing confirmed this gap — the model performed admirably in simulated conversations but faltered in real-world scenarios.</p><p>The thing is, synthetic data often looks too polished and artificial — far from how people actually communicate. To prevent our model from sounding unnatural, we again leveraged human expertise. What we’ve done again was introducing actual customer dialogues from our banking partner not just to our pre-train, but also to our fine-tuning process, carefully anonymized and reviewed by both banking and language experts. Banking professionals ensured the accuracy of financial information, while language specialists maintained conversational naturalness. It’s worth saying that even without real data, we could have managed to fix this “unnaturalness” through post-training alignment, a process of fine-tuning AI models to make their outputs more natural and human-like, essentially teaching the model to communicate more authentically by understanding nuanced conversational contexts. However, that would have required more time and money.</p><p><strong>Are we on the right track?</strong></p><p>To understand whether our training process was heading in the right direction, we evaluated multiple aspects of our model: instruction comprehension, response generation, factual accuracy, grammar, adherence to Uzbek language, context handling, and safety compliance.</p><p>We started with industry-standard English datasets (MMLU, Truthful-QA, Halu-Eval), removing irrelevant domains like medicine, law, and history. We categorized data based on input length (short questions vs. long instructions with context) and output length requirements (from single words to hundreds of words). All language-specific samples were removed and translated into Uzbek, with special techniques applied to reduce translation noise.</p><p>Testing the model presented significant challenges. We needed to evaluate numerous aspects simultaneously, and complete test sets contained thousands or tens of thousands of examples. Running full evaluations could consume an entire day of processing time. To address this, we implemented strategic random stratified sampling from different test sets, creating a manageable evaluation corpus of approximately 2,000 samples (excluding safety evaluations).</p><p>To validate our approach, we assessed the correlation between our aggregated metrics and established Elo scores from LM-Arena using known English-language models. Finding a statistically significant (p-value &lt; 0.01) positive correlation confirmed our methodology was sound. Our working assumption was that if this metric approach proved effective for English models, it would, with appropriate adjustments, work similarly for our Uzbek language models.</p><h3>Field Test: A Bilingual Banking Assistant in Telegram</h3><p>Practice makes perfect, and training is nothing without feedback collection, so even after excessive fine-tuning we needed some place to test it all on real users before any kind of full scale product shipping. In other words, we needed a functional prototype into testing as quickly as possible to establish a managed development cycle (deployment → feedback collection → refinement → redeployment) with short iterations. Without an early prototype, we risked making insufficient improvements.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/779/0*azTC_RrRYPEiyF4s" /><figcaption>Expected relative increase in prompt injection response relevance. Comparison between a fine-tuned model, a fine-tuned + alignment model, and a reference model (GPT-4o).</figcaption></figure><p>Telegram proved to be an ideal platform for rapid chatbot prototyping. Its asynchronous API, support for integrating various models and scenarios, availability across web, desktop, iOS/Android platforms, and most importantly its widespread popularity in Uzbekistan, made it the clear choice for our needs.</p><p>After weeks of fine-tuning and testing, we successfully deployed our bilingual LLM as a Telegram bot for our banking client. The bot effectively handled inquiries in both Uzbek and Russian, guiding users (bank employers) through banking processes like loan applications and account management, and is now ready for implementation in other banking products.</p><p>A significant advantage of this approach was the ability to involve native Uzbek speakers from the client’s side in testing, despite our development team lacking native speakers. This collaboration quickly revealed some initial issues — for instance, early versions of the LLM sometimes confused numerals. We identified and fixed these problems promptly thanks to the prototype.</p><p>One of the most interesting outcomes was the model’s ability to handle language-switching between Uzbek written in Cyrillic and Latin scripts. For example, a user might type in Uzbek using Cyrillic, and the bot would respond in Uzbek using Latin script, with the conversation continuing seamlessly. This unexpected capability demonstrated the model’s adaptability and enhanced its practical value for users.</p><h3>Lessons Learned</h3><h4>1. The Challenge of Low-Resource Languages</h4><p>Uzbek is a low-resource language, and adapting LLMs for such languages requires creativity. We had to rely on a mix of automated translation, native texts, and synthetic data to achieve desired results.</p><h4>2. The Importance of Real-World Data</h4><p>Synthetic data is a great starting point, but real-world interactions are crucial for fine-tuning. Without them, the model may struggle to handle the nuances of human conversation.</p><h4><strong>3. Nothing beats real</strong></h4><p>Whenever possible, quickly implement things to any safe, but real environment (like a telegram bot) to gather as much feedback as possible — nothing beats invaluable real user experience.</p><h4><strong>4. The simpler the better</strong></h4><p>The simplest approaches, which involved training the fewest parameters possible, worked best for us. We found that the more parameters we tried to fine-tune, the more pronounced the negative effects of noise in the data became. This reinforced the idea that, especially in narrow-domain tasks, less can often be more — keeping the model focused and minimizing unnecessary adjustments led to better performance and more reliable results.</p><p>…and just in case you’re a business person reading this blog here’s one more bonus lesson learned:</p><h4><strong>5. Regional Leadership Demands Technological Sovereignty</strong></h4><p>TBC Uzbekistan’s approach offers a valuable insight for businesses across emerging markets: true market leadership requires technological sovereignty. By investing in proprietary AI solutions tailored to local languages and cultural contexts rather than relying on generic global solutions, TBC established competitive advantages that cannot be easily replicated. The commitment to building internal AI expertise while strategically leveraging external partnerships demonstrates how companies can maintain independence while accelerating innovation.</p><h3><strong>Business implementations and what’s next</strong></h3><p>Together with TBC Uzbekistan we implemented trained LLMs in Sales Pro — a “digital employee” that communicates with potential customers and offers them bank products, such as loans. This implementation fully leveraged both our advanced language models and the GOAL framework as the controlling architecture.</p><p>When selling, the bot’s task isn’t just to convey information to the customer, but to engage them, discuss their needs, articulate how a loan can address those needs, handle objections, and gently guide the customer toward completing an application. Thanks to TBC Uzbekistan’s substantial investment in AI capabilities and the collaborative effort with Aiphoria providing methodology and platform support, this ambitious region-wide project achieved exceptional results.</p><p>As a result, TBC Uzbekistan opened a new sales channel through a Telegram bot. After just one month of operation, it was already showing impressive results — conversion rates to banking product applications (such as loans) were <strong>one-third higher</strong>, and the acquired users were <strong>twice as engaged</strong> compared to other acquisition channels, including those requiring human involvement (like telemarketing). The difference is that in this case no staff is required and the automated product is just as good and even better at convincing people to buy things.</p><p>The next goal in TBC Uzbekistan’s AI roadmap is cold calling. This traditionally requires significant employee time and energy, particularly when reaching out to completely unfamiliar prospects. The “digital salesperson” complements human efforts by handling initial outreach without experiencing fatigue from rejections, maintaining consistent communication quality throughout all interactions. This approach allows for successful customer conversions even with challenging audiences while reducing costs and freeing human staff to focus on more complex, relationship-building aspects of sales that benefit from a personal touch.</p><p>***</p><p>In the next story we’ll cover the process behind making LLM answers as fast as possible, which is crucial in phone call cases. There’s a story to be told of adapting both LLM and ASR/TTS parts of our system to make it so. Stay tuned!</p><p><em>Written by: </em><strong>Aiphoria NLP team</strong> (Ivan Pastukhov, Nikita Pustovoytov, Anton Voronov, Denis Kirianov)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=663e6b3891b8" width="1" height="1" alt=""><hr><p><a href="https://medium.com/aiphoria/when-languages-lack-data-making-open-source-llms-work-for-banking-in-uzbekistan-663e6b3891b8">When Languages Lack Data: Making Open-Source LLMs Work for Banking in Uzbekistan</a> was originally published in <a href="https://medium.com/aiphoria">Aiphoria</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI agents in an overlooked world. A story of developing modern speech tech in Uzbekistan.]]></title>
            <link>https://medium.com/aiphoria/ai-agents-in-uzbekistan-cad2f5b3c1b3?source=rss-7c54839448b6------2</link>
            <guid isPermaLink="false">https://medium.com/p/cad2f5b3c1b3</guid>
            <category><![CDATA[speech-synthesis]]></category>
            <category><![CDATA[speech-recognition]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[banking-technology]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Aiphoria Team]]></dc:creator>
            <pubDate>Fri, 07 Feb 2025 15:03:37 GMT</pubDate>
            <atom:updated>2025-02-07T15:03:37.168Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/proxy/1*IiVZemIIrydTgo1GDWnD1w.jpeg" /></figure><p>Hello everyone! We are <a href="https://aiphoria.ai/">Aiphoria</a> — a company developing solutions that automate routine employee work across organizations. There’s a lot of buzz around AI Agents lately, with some calling it one of the <a href="https://www.gartner.com/en/documents/5850847">major trends</a> for 2025. We call our virtual employees “AI Pros” and develop them as turnkey solutions for various business needs — from customer support to recruitment and much more.</p><p>We have developed our own next-gen NLP platform (Aiphoria Pro Platform) that makes it incredibly easy to integrate AI Pros into almost any business, even if its operations are not just text- but also voice-centric. It’s adaptable to any specific knowledge domain and language, features an open architecture, supports on-premise deployment, and most importantly — our pros can achieve set goals thanks to our flexible GOAL architecture (GOAL stands for Goal-Oriented Architecture for LLM). This means we ensure that our clients’ customers are guided not just through a predetermined rule-based scenario but toward a specific objective, which can range from selling something to solving a particular issue in tech support scenarios. All of this is thanks to advanced implementation of GenAI, especially LLM’s conversational capabilities. We’ll definitely cover this approach in more detail in our future posts.</p><p>You’ve probably read numerous stories about how such things work for the English language, so we won’t dwell on the details here. Instead, we wanted to dedicate this first post to a challenge we encountered in one of our cases, which is quite representative of the current state of AI in general.</p><p>As we’ve mentioned before, we can help with not just text-based scenarios but also with voice-centric ones — we find the possibilities of current LLMs for these especially exciting. While speech tech in English-speaking countries is nothing new, trying to achieve the same for, say, Central Asian countries is a challenge of a different magnitude. These countries are often overlooked in terms of AI technologies, despite being developed nations with unique cultures and massive digitized businesses across various sectors, which makes them data-rich environments. Today we’ll share one such case study — a collaborative effort with TBC Uzbekistan, the leading digital banking ecosystem in Central Asia, where we worked together to build and implement full-fledged AI agents. Despite our team’s experience working in big companies on various challenging products, what we’ve achieved together in less than a year will always have a special place in our hearts and makes us really proud; we truly consider it a remarkable tech achievement by any means.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*D0oibNw-PbfLL3HM" /><figcaption><em>Image credit: TBC Uzbekistan</em></figcaption></figure><p><strong>The Challenge</strong></p><p>Let’s summarize the challenges we tackled together. While we brought the technical expertise, TBC Uzbekistan brought deep banking knowledge and a clear vision of what they wanted to achieve: they needed virtual operators to ease the workload of their call center across several scenarios — including support, payment reminders, sales, and potentially much more. They played a key role in shaping these solutions, providing quick feedback and valuable insights from their digital banking operations. Given the sensitive nature of banking data, data confidentiality was paramount from day one, which made our platform on-premise capabilities a perfect fit for this partnership.</p><p>The challenge was to make everything work for the not very common Uzbek language. What does it take for us to localize Aiphoria Pros for a new language? We need to create speech synthesis and recognition, support NLP models, and train LLM to support this language with decent quality. Let’s start with speech synthesis and recognition first, two related but uniquely challenging tasks. But before that — some geographical and cultural context is probably needed.</p><p>Uzbekistan is a vast and beautiful country in Central Asia bordering Tajikistan, Kyrgyzstan, and Kazakhstan. It is also the most populous country in the region, with more than 37 million people. It is a former USSR country with several languages in use, which are often mixed together. This is especially relevant in border regions, where people might speak Tajik or Karakalpak — you probably haven’t heard of the latter, but it has 2 million native speakers.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*2dW5pZaLLPYCv_pi" /><figcaption><em>Uzbekistan is a diverse country in more than one way as evident by these photos taken from </em><a href="https://ru.wikipedia.org/wiki/%D0%A2%D0%B0%D1%88%D0%BA%D0%B5%D0%BD%D1%82-%D0%A1%D0%B8%D1%82%D0%B8"><em>various sources</em></a><em>, including our founder’s iPhone</em></figcaption></figure><p>Add to this various dialects, such as Tashkent, Surkash, Jizzakh, Namangan, and Khorezm dialects, each sounding quite different. Another interesting and important fact is the country’s relatively recent transition from Cyrillic to official Latin script, which many people still might not know. Additionally, a lot of folks often mix languages in their day-to-day communication, speaking several languages even in one sentence. Consider all this — and you get an extremely interesting challenge for any ML engineer trying to build speech tech here. And let’s not forget the traditional challenges of speech analysis from phone call data in most countries: poor call quality, calls from subway stations, traffic jams, villages with bad mobile connection etc.</p><p>Blog readers might not know Uzbek and Russian languages, so let’s try to imagine how this would sound with more familiar Spanish and English languages. Imagine receiving a call where a barely audible voice says something like: “Hey there, es un bank? My card no trabajo and I’m in deep desierto with connection débil”. Only it’s not in English and Spanish, it’s in Uzbek with a mix of Russian, and it’s more or less regular speech that you can hear on the streets.</p><p><strong>The state of Uzbek Automatic Speech Recognition</strong></p><p>Before building our own solution for Uzbek ASR, we needed to understand what already existed in the market. It turned out that several solutions already exist; however, they all have various issues that make them unfit for our purpose.</p><p>Evaluation of quality for such products is usually done using SER/CER (sentence/character) metrics or WER (Word Error Rate) metric. We’ve used the latter, which is a standard, widespread metric for measuring speech recognition quality that shows the percentage of errors per given number of words. The lower the percentage, the better.</p><p>On paper, all existing solutions show good results — for instance, “META seamless_communication” showed 17.33% WER on the FLEURS dataset. However, when faced with real raw data from TBC Uzbekistan’s call center, with usual phone call quality and specifics of banking vocabulary, the actual quality of existing solutions turned out to be at best around 40% WER. To roughly explain what this means, you can imagine that in an Uzbek sentence of 10 words, existing solutions correctly recognize only 6. Unfortunately, this is an unacceptably low result.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/966/1*SM629EtAg5AVkRk_-WJGFg.png" /></figure><p>Without proper ASR, AI agents that have to work with voice channel are not possible, as this is the base of it all. Too many mistakes in recognition will eventually lead to LLMs not getting what the problem in question is and simply not working properly. That’s why we had no choice but to start building a new speech recognition system specifically for TBC Uzbekistan.</p><p><strong>Building our own ASR</strong></p><p>The foundation of any recognition system is well-labeled datasets. Fortunately, the recent trend is global open-source solutions, both from large companies like Meta and startups. Sharing training data is beneficial for everyone because it’s impossible to achieve good results quickly in isolation — constant retraining is needed, and each new project using open data has the potential to improve that data’s quality. This is relevant even with complex Foundation Models, as evidenced by <a href="https://www.statista.com/statistics/1465314/foundation-models-by-access-worldwide/">Stanford’s research</a> here:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*cbJDTPFvRyYUVQm8" /><figcaption><em>Artificial intelligence foundation models worldwide, 2021 to 2023, by access type (Statista)</em></figcaption></figure><p>This is also relevant for datasets in different languages. In the case of Uzbek, the team from Uzbek Voice worked in partnership with Uzbekistan’s IT Park and decided to make their impressive 1000+ hours Uzbek speech dataset publicly available. There are also other examples like Mozilla Common Voice, which is much smaller, and the multi-language FLEURS dataset with just a dozen hours of Uzbek, to name a few.</p><p>Unfortunately, this is not enough to build a functioning ASR for our purposes. Big datasets are never perfect, and there are usually many mistakes in transcriptions. For instance, while analyzing one of these datasets, we came across historical Uzbek verses transcribed with obsolete terminology that proved impossible to identify within the text corpus. And this is just one example. Most importantly, however, these datasets lack necessary data from conversations with banks — all the specifics of the banking lexicon, certain terms, and definitions. That’s why the call center data from TBC Uzbekistan, which we’ve mentioned before, was crucial.</p><p>So we have this labeled but not perfect data from public datasets and anonymized unlabeled crucial data from the TBC Uzbekistan call center. Sounds like quite a task for a labeling team! To understand its importance better, let’s first understand what data labeling essentially is.</p><p>Data labeling is another foundation of effective ASR systems. The process starts with creating high-quality audio-text pairs, where human experts carefully transcribe speech recordings, noting everything from words to emotional tones. Think of it as building a comprehensive translation dictionary between spoken and written language.</p><p>But raw transcriptions aren’t enough. The next crucial step is normalization — transforming these transcripts into a standardized format. This cleanup process ensures consistency across different speakers and accents, making it easier for AI models to learn patterns and generate accurate transcriptions. It’s like establishing a universal grammar that helps the system understand and process speech more effectively, regardless of who’s speaking.</p><p>Remember how we mentioned earlier that Uzbekistan switched from Cyrillic to Latin script? And how there is a mix of different languages and dialects? All of this made building a data labeling team a non-trivial task. For example, there aren’t many people who know the official Uzbek language (i.e., Latin writing). Therefore, we first had to build expertise internally in the bank and bring in linguists who could explain the nuances of grammar and dialects.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*N0XtAv_3w1d1jLNj" /></figure><p><em>To this day there’s a mix of scripts and languages even on the signs of street shop vendors © Yandex Maps</em></p><p>Normalizing both existing labeled datasets and our annotators’ output was quite the adventure. We had to standardize everything from various spellings of names and places to some rather peculiar linguistic cases.</p><p>Let us share some examples. In the official Uzbek language, the letters ‘h’ and ‘x’ create an interesting confusion — the ‘x’ isn’t pronounced like in English but almost identically to ‘h’, just slightly firmer. Given the recent transition from Cyrillic to Latin alphabet, this led to multiple transcription variants. Take the word “xo’p” (equivalent to “OK” in English) — we’ve seen it written as “hop”, “ho’p”, “hup”, and “xop”. While these variations are commonly used, even by our annotators, they’re technically incorrect. The word is so prevalent in the language that sometimes almost an entire conversation can consist of “xo’p”. This would occasionally lead to our model “getting creative” and translating some similarly sounding names and surnames as “Xop Xopich” (the English equivalent would be something like “Mr. Okay Okayson”). Without standardizing these variations, we’d risk sounding illiterate.</p><p>Then there’s the very common phrase “assalomu alaykum,” which means “peace be upon you” but is basically used as “hello”. We used the correct form in the previous sentence, but there are many examples of incorrect spellings like “assalomy,” “assalamu,” “alekum,” “aleykum” and many others with different combinations — all of these need to be normalized. Again, it’s not that people are illiterate; it’s all part of a recent transition from the Cyrillic to Latin alphabet — people are not yet used to the different letters.</p><p>Another challenging case involved the ‘-lik’ suffix with numerals. Even our linguistic expert had to consult external colleagues to crack this one. We discovered that ‘-lik’ should only be used when a numeral represents something collective or characteristic of a group — like area codes for phone numbers — but not with card numbers. Getting these subtle distinctions right was crucial; otherwise, our model could misinterpret user inputs and generate incorrect responses.</p><p>In total, we had to make over 7,000 rules for such normalizations.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/950/0*ub94hkBqat25_wbh" /><figcaption><em>Just to name a few</em></figcaption></figure><p>Another challenge was learning to identify specific vocabulary, as not all annotators understand certain banking nuances. For instance, team members didn’t understand what “KATM” was, despite its frequent mention in calls — it’s simply the equivalent of a credit bureau. In such cases, we often encounter “acoustic illusion” — when all five people are absolutely certain they’ve correctly transcribed what was said, but none of them are right. This is partially solved by properly selected datasets — we based ours on customer-operator conversation topics and created a similar domain for the training corpus. Part of the internal dataset was collected with help from call center operators, whom we asked to record phrases containing useful vocabulary and internal bank abbreviations. In some cases, we had to bring in people from other departments to help resolve disputed cases.</p><p>As a result, our data labelling team became so proficient that some team members started receiving comments from friends about how grammatically correct their chat messages had become.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/781/0*Hw9gKXQubUCHIL3w" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/781/1*-CIZIdUDfB9r0PW5OjW_qg.jpeg" /><figcaption><em>A real </em>conversation where a member of our team politely asks to send photos and gets a response that he talks ‘like a teacher,’ meaning he speaks very ‘properly.’</figcaption></figure><p>Now we have sufficient well-annotated data and have created our own working ASR — likely the world’s best for the Uzbek language. This is confirmed by subsequent production tests and specific results on public datasets.</p><p><strong>Speech Synthesis</strong></p><p>English-speaking audiences are no strangers to good speech synthesis — from voice assistants to navigation and customer support in various services. The situation is also good with other popular languages, at least with most European ones. However, there was no on-premise deployable model for the Uzbek language with good enough quality in both ASR and TTS to address TBC use cases (bilingual and tonally diverse bots mimicking humans in phone calls).</p><p>Typically, people in Uzbekistan hear what is called unit selection synthesis — where pre-recorded phrases change depending on the situation. For example, in a pre-recorded sentence, a number, street name, or name might change, but the sentence structure remains unchanged. It often sounds quite robotic and unconvincing. While all speech synthesis once worked this way, for many languages there is now synthesis that is sometimes barely distinguishable from human speech. Why should Uzbek be any different? That’s why we set out to create exactly this kind of quality with full-fledged neural network-based synthesis where you can synthesize any word from scratch — not just a pre-recorded one. It solves the bank’s tasks much better — both as human support and for other scenarios like payment reminders.</p><p>At its core, a modern TTS pipeline consists of three essential components. The first is the NLP component, which handles all text processing and produces “clean” text input, properly formatted without numbers or special punctuation marks. It is quite a challenge by itself, especially in languages with complex declension of words like in Russian and Uzbek — a simple “4%” is ofter written like this, but could sound very different depending on the case. The second component is the acoustic model, which generates mel spectrograms from the processed text and forms the heart of the synthesis process, trained on numerous hours of audio-text pairs. Finally, there’s the vocoder that transforms mel spectrograms into waveforms that we can hear, typically outputting .wav files or sound streams.</p><p>None of the existing speech synthesis solutions suited us for several reasons: we’re building a bilingual multi-tonal solution that we want to work autonomously on the bank’s servers. Bilingual means that TTS should synthesize the same voice in both languages in order to be able to communicate with all people in the country — both Uzbek and Russian speaking; even if they switch languages during a conversation, we can adapt and answer accordingly without changing voices. While we’re also dealing with potentially difficult situations, be it support or payment reminder scenarios, the voice should be humane and have a range of tones suited for various cases.</p><p>One of our biggest initial challenges while building TTS was the lack of specific open-source datasets for Uzbek needed to make a solution fit for banking with all its vocabulary and situations. This meant collecting and preparing vast amounts of text data from various sources, including internal anonymized (non-sensitive) banking info, financial news, and general content like sports coverage and fairy tales. A proper TTS dataset should also be representative across multiple dimensions: high-frequency words, phoneme combinations, and various other linguistic patterns. We carefully curated our dataset with all these factors in mind.</p><p>Our team has also developed several other crucial components from scratch, including a custom normalizer for Uzbek that can handle specific cases like ordinal numbers, converting them into their written unified form. We also created an accentuation system for Uzbek words.</p><p>The other challenge was that in order to train a TTS model, we needed many hours of recording Uzbek and Russian speech, spoken by the same person — and like before, this is not only regular speech; it should also include specific banking terms and other financial characteristics. Thankfully, having already collected the needed amounts of text for a well-balanced dataset, we found a wonderful young woman who is an extremely professional voice actor speaking both Uzbek and Russian excellently.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*3RBo1FmgDzPrOhGV-_WQKQ.jpeg" /><figcaption>Work in progress</figcaption></figure><p>As we’ve mentioned before, just one voice tone simply doesn’t cut it when dealing with the variety of situations in banking. A friendly reminder about an upcoming payment and an urgent notice require completely different approaches. The tone of voice (TOV) can make or break these interactions, potentially influencing the customer’s response and the conversation’s outcome. That’s why we’ve developed three distinct voice tones:</p><ul><li>Friendly: for support and routine interactions</li><li>Neutral: for standard communications</li><li>Strict: for urgent matters</li></ul><p>Behind the scenes, this meant our voice talent had to record extensive amounts of text in these different emotional states and in two languages. We didn’t just wing it — the heads of TBC Uzbekistan’s respective departments carefully supervised the recording process to ensure each tone perfectly matched their communication needs and didn’t get mixed with another.</p><p>From an ML perspective, our approach was methodical:</p><p>1. First, we developed a base model using our complete voice dataset</p><p>2. Then, we fine-tuned specific emotional variations using carefully curated subsets</p><p>3. Finally, we excluded inappropriate tones based on the intended use case</p><p>We added an extra layer of sophistication by incorporating tongue twisters into our training data. These traditional sayings and proverbs contain words with unique phoneme combinations that rarely appear in everyday speech. This proved to be a clever insurance policy — if our model ever encounters these unusual letter combinations in the wild, it already knows how to pronounce them correctly.</p><p>The results? A versatile TTS system that speaks the language of banking in all the right tones, capable of handling even the most unusual pronunciations.</p><p>Few examples of Uzbek and Russian synthesis of a phrase “AI will make our lives better”:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fw.soundcloud.com%2Fplayer%2F%3Furl%3Dhttps%253A%252F%252Fapi.soundcloud.com%252Ftracks%252F2022851296%26show_artwork%3Dtrue&amp;display_name=SoundCloud&amp;url=https%3A%2F%2Fsoundcloud.com%2Fsomebody-74110847%2Fsynthesis-example%3Futm_source%3Dclipboard%26utm_medium%3Dtext%26utm_campaign%3Dsocial_sharing%26si%3Dc474af10c05542bd8e186444ccb4cb64&amp;image=https%3A%2F%2Fsoundcloud.com%2Fimages%2Ffb_placeholder.png&amp;type=text%2Fhtml&amp;schema=soundcloud" width="800" height="166" frameborder="0" scrolling="no"><a href="https://medium.com/media/887306f8261f802e0cf225dce298a423/href">https://medium.com/media/887306f8261f802e0cf225dce298a423/href</a></iframe><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fw.soundcloud.com%2Fplayer%2F%3Furl%3Dhttps%253A%252F%252Fapi.soundcloud.com%252Ftracks%252F2022851912%26show_artwork%3Dtrue&amp;display_name=SoundCloud&amp;url=https%3A%2F%2Fsoundcloud.com%2Fsomebody-74110847%2Fsynthesis-example-rus%3Futm_source%3Dclipboard%26utm_medium%3Dtext%26utm_campaign%3Dsocial_sharing%26si%3D0ad7c3066fc74db9bc29fb8cc0eb34ee&amp;image=https%3A%2F%2Fsoundcloud.com%2Fimages%2Ffb_placeholder.png&amp;type=text%2Fhtml&amp;schema=soundcloud" width="800" height="166" frameborder="0" scrolling="no"><a href="https://medium.com/media/6f21558424db3d390b9c5318339b0c07/href">https://medium.com/media/6f21558424db3d390b9c5318339b0c07/href</a></iframe><p>Working with a professional voice actor in a studio setting brought its own set of interesting challenges. Ironically, her high level of professionalism sometimes worked against us — we occasionally needed to ask her to speak more naturally rather than using her perfect professional diction. So we also needed to monitor the voice to ensure it doesn’t strain and that intonations don’t change due to fatigue.</p><p>Additionally, any recording process for speech synthesis purposes is teamwork between the sound engineer and the voice talent. In our case, the sound engineer not only monitors the quality of the recording process but also checks the text’s correctness. But not everything is quite easy — not in our case! Going back to the three-alphabet problem: it turned out that the engineer couldn’t read Cyrillic but understood Latin script. The voice talent, on the contrary, could read Cyrillic perfectly but couldn’t read Latin script at all. This meant we had to prepare two versions of the texts for recording — one in Cyrillic and one in Latin script, so everyone could understand and monitor potential errors. The result was 20 clean hours of recorded text without pauses, which meant more than 100 hours in the studio.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F6ySusQITaT0%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fshorts%2F6ySusQITaT0%3Ffeature%3Dshare&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F6ySusQITaT0%2Fhq2.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/544a6479333947b09353ad22c4987b0e/href">https://medium.com/media/544a6479333947b09353ad22c4987b0e/href</a></iframe><p>Just like department heads supervised our voice recordings, we needed people who truly understood the nuances of customer communication to evaluate our resulting TTS models. External assessors wouldn’t cut it — they simply wouldn’t know how support agents actually talk to customers.</p><p>Our solution? We built our own assessment system from scratch, recruiting call center veterans who lived and breathed customer service every day. These pros rated each model iteration on a 1–5 scale, focusing on natural flow and appropriate tone of voice. As a side note, talking about new professions created by AI: in our opinion, some of these people are now perfect voice coaches and, if they wish, they can work in this booming industry instead of their old jobs in support.</p><p>While we’ve achieved significant success with the acoustic model, we’re currently working on fine-tuning an open-source vocoder to better match our speaker’s characteristics. The substantial dataset we’ve created for a single-speaker bilingual system provides an excellent foundation for future improvements.</p><p>After half a year of dedicated collaboration, our joint efforts with TBC Uzbekistan yielded something remarkable — together we brought to life what’s likely the most advanced speech tech solution for a country of almost 40 million people. The solution was shaped not just by our technical expertise, but also by TBC’s deep understanding of their customers’ needs and their exceptional readiness to innovate and implement new features. Their ability to quickly test, validate, and roll out new capabilities contributed heavily to making it possible to achieve in months what typically takes years.</p><p>From a business perspective, the partnership has already proven transformative: as of early 2025, AI-powered agents handled over 40% of call reminders for early-stage delinquency loans, delivering efficiency gains while maintaining a human-like conversational experience that most customers couldn’t distinguish from real people. This innovative implementation of AI has significantly boosted operational efficiency, being 10 times more efficient than humans. Encouraged by these results and maintaining their characteristic momentum, TBC is taking an active role in expanding this technology across other business functions, continuously working with us to identify new opportunities and use cases.</p><p>What started as a solution for Uzbekistan revealed a much bigger opportunity. We discovered that many regions face similar challenges — places where global speech recognition solutions fall short of local needs. Even Spanish-speaking countries, despite having access to major providers like Google, struggle with regional variations; Argentine Spanish, for instance, differs significantly from the standard Spanish these solutions support.</p><p>We believe that our methodology has become a product in itself, enabling organizations to build locally-optimized speech technology that truly understands their customers. We’re offering proven expertise in helping companies develop custom solutions using their own data. This foundation then unlocks the full potential of our platform, allowing companies to deploy AI agents across various parts of the business.</p><p>Stay tuned for the equally interesting next part of our story — the building of an LLM tailor-made for banking needs and making it work in pair with our ASR and TTS.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cad2f5b3c1b3" width="1" height="1" alt=""><hr><p><a href="https://medium.com/aiphoria/ai-agents-in-uzbekistan-cad2f5b3c1b3">AI agents in an overlooked world. A story of developing modern speech tech in Uzbekistan.</a> was originally published in <a href="https://medium.com/aiphoria">Aiphoria</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>