-
151
Projects submitted
-
109
Published projects
-
47
Comments left
-
66
Number of votes
-
52
days until voting ends
Omnichannel contact center with AI and Uzbek speech recognition
-
Project manager on customer side
Shukhrat Agzamov
Uztelecom
Head of R&D and Innovations
-
Category
-
Nomination
-
Goals
Transform the customer service of the national telecom operator Uzbektelecom (13.65 million subscribers) by unifying all voice and digital communication channels on a single omnichannel platform with fully localized artificial intelligence, Uzbek speech recognition, and automated analytics for 100% of requests.
-
Timeline
march, 2024 — november, 2025 -
Project completion year
2025
-
Project scale
5000 human-hours -
Results
Proprietary Uzbek ASR Deployed: Trained in-house model on ~625h of audio (~150h of real contact center calls), covering regional dialects and telecom terms.
High Accuracy (~10% WER): Achieved ~10% Word Error Rate on noisy real-world telephone calls, outperforming off-the-shelf cloud engines.
Omnichannel Consolidation: Unified 4 core channels (Voice, Telegram, Facebook, Website) for 13.65M subscribers with automated operator workload balancing.
100% Automated AI Call Analytics: Replaced manual sampling with AI-generated call summaries, key takeaways, topic tagging, and script compliance scoring.
100% On-Premise Sovereignty: Deployed locally on Uztelecom’s Tier III GPU servers (ISO/IEC 27001:2022, PCI DSS); zero subscriber audio leaves the corporate contour.
B2C Monetization ("Mobile Secretary"): Launched a commercial AI assistant (10,500 UZS/mo via MyUTC & USSD *6525#) with live UZ/RU dialogue, spam filtering, and text transcripts.
-
Project uniqueness
Unlike contact centers that rely on third-party public cloud APIs, JSC Uztelecom built a 100% sovereign, on-premise Speech & Generative AI ecosystem tailored specifically for the Uzbek language and its regional dialects. Trained on ~625 hours of audio (~150 hours of real noisy contact center calls), our proprietary ASR model achieves an industry-leading ~10% Word Error Rate (WER) on live telephone lines with Uzbek/Russian code-switching. Running entirely on Uztelecom’s own Tier III NVIDIA GPU infrastructure, the platform guarantees that subscriber voice and personal data never leave the corporate security contour. Uniquely, the same core speech AI architecture simultaneously powers internal enterprise operations (omnichannel routing, 100% LLM call summarization, and automated script scoring) and external B2C monetization via the "Mobile Secretary" virtual assistant for both smartphones and feature phones. -
Software used
Omnichannel & Telephony Core: In-house Omnichannel Contact Center Platform & Intelligent Operator Workload Management System (JSC Uztelecom / Telecom-Soft), SIP/VoIP trunking, WebRTC, Telegram Bot API, Meta (Facebook) Graph API, and USSD gateway (*6525#).
Speech AI & Generative AI Stack: Proprietary fine-tuned Uzbek/Russian Automatic Speech Recognition (ASR) acoustic and language models (trained on ~625h dataset), Voice Activity Detection (VAD) & noise suppression pipelines, fine-tuned Large Language Models (LLM) and NLP pipelines for automated call summarization, topic classification, sentiment analysis, and script compliance scoring, plus the SI-Sufllyor real-time agent-assist module.
AI Infrastructure & Orchestration: On-premise NVIDIA GPU server clusters hosted in JSC Uztelecom Tier III Data Centers (UzCloud), Kubernetes container orchestration, Run:ai GPU resource scheduling, KVM virtualization, high-speed NVMe storage, and Infrastructure-as-Code (Terraform, Ansible). -
Solution from the Global CIO catalog
The project does not use solutions from the Global CIO catalog
-
Project implementation complexity
-
Zero Ready-Made Uzbek Speech Datasets: Off-the-shelf ASR engines failed on regional Uzbek dialects, informal speech, and 8 kHz line noise; our team had to manually curate, clean, and annotate ~625 hours of audio (~150 hours of live calls) from scratch.
-
Bilingual Code-Switching & Telecom Jargon: Subscribers frequently mix Uzbek and Russian mid-sentence while using technical telecom terms (tariffs, billing, GPON/5G), requiring custom tokenizer and acoustic adaptation.
-
Strict On-Premise & Low-Latency Constraints: Under Uzbekistan’s data localization laws (ZRU-547), external cloud AI APIs could not be used. Sizing and orchestrating high-density local GPU inference to process thousands of concurrent voice streams and LLM summaries with minimal latency was a major engineering challenge.
-
Legacy & Multi-Channel Integration: Seamlessly unifying legacy SIP/VoIP telephony, social messengers, CRM/billing, and real-time workload balancing without disrupting 24/7 support for 13.65M subscribers
-
-
Description
Initial Situation: As the national telecommunications operator of Uzbekistan serving 13.65 million subscribers, JSC Uztelecom processes over 200,000 customer inquiries daily across more than 1,500 contact center agents. Previously, voice calls and digital channels (Telegram, Facebook, website) operated in fragmented silos without unified customer history. Quality assurance relied on manual supervisory listening of less than 2–3% of calls. Furthermore, global off-the-shelf cloud speech recognition (ASR) engines demonstrated poor accuracy (>35–40% word error rate) on conversational Uzbek, regional dialects, Russian-Uzbek code-switching, and noisy telephone lines, while strict national data localization and telecom secrecy laws prohibited transmitting subscriber voice recordings to external public clouds.
Implemented Solution: JSC Uztelecom, together with its IT subsidiary Telecom-Soft, designed and deployed a unified, AI-powered Omnichannel Contact Center platform running 100% on-premise. The solution integrates four core communication channels (Voice/SIP, Telegram, Facebook, and Website live chat) into a single agent workspace coupled with an intelligent operator workload management system. At the core of the platform is a proprietary Uzbek ASR model trained in-house and an LLM/NLP-based post-call analytics and real-time agent-assist module (SI-Sufllyor) that transcribes 100% of calls, generates structured call summaries, extracts key highlights, classifies inquiry topics, and automatically scores agent script compliance. In addition, the same speech AI engine was commercialized into the B2C "Mobile Secretary" virtual voice assistant (accessible via the MyUTC app and USSD
*6525#).Implementation Specifics: To overcome the lack of public datasets for conversational Uzbek, the R&D and AI team collected, anonymized, and annotated ~625 hours of domain-specific audio, including ~150 hours of real contact center dialogues across all regional dialects of Uzbekistan. Acoustic and language models were fine-tuned specifically for 8 kHz noisy telephony channels and bilingual (Uzbek/Russian) telecom terminology. The entire AI inference pipeline was deployed locally on Uztelecom’s own NVIDIA GPU clusters inside Uptime Institute Tier III certified data centers (compliant with ISO/IEC 27001:2022 and PCI DSS), ensuring low-latency real-time processing and complete data sovereignty.
Changes After Implementation: Customer service shifted from fragmented channel handling to a single omnichannel workspace with end-to-end subscriber context and automated queue balancing. Word Error Rate (WER) on noisy real-world Uzbek calls dropped to ~10%. Quality assurance transitioned from 2–3% manual sampling to 100% automated AI evaluation of every call, freeing supervisors from routine listening and reducing post-call wrap-up time for agents via automated LLM call summarization. Furthermore, the internal R&D investment was transformed into a new revenue stream by launching the B2C "Mobile Secretary" service (10,500 UZS/month), which answers missed calls in Uzbek and Russian, blocks spam, and sends text transcripts to subscribers.
-
Project geography
Call Center of "Uztelecom" JSC
-
Additional presentations
-
Customer
Uztelecom
-
IT provider
Telecom Soft
Only authorized users can comment.
We suggest you log in or register.