←All articles
ИИ-Агенты & Voice AI2026-11-19 · 9 min

Разработка голосовых ИИ-агентов и AI-колл-центров в Казахстане: синтез казахской и русской речи, SIP-телефония, интеграция с Kaspi QR и CRM

Экспертное руководство по внедрению голосовых ИИ-агентов нового поколения (Voice AI) для бизнеса в Казахстане (Алматы, Астана, Шымкент) и на международном рынке. Стек: WebRTC/SIP, Silero VAD, Whisper с поддержкой код-свитчинга, Cartesia/ElevenLabs TTS, генерация счетов Kaspi Pay и синхронизация с AmoCRM/Битрикс24.

The Crisis of Traditional Call Centers in 2026

Phone communication remains the highest-converting touchpoint for B2B and B2C sales and urgent customer inquiries. Yet companies in Almaty, Astana, Shymkent and global tech startups worldwide face persistent operational bottlenecks:

1

Losing 40–50% of Inbound Leads: During peak hours and off-business times, callers encounter long ring tones or hold queues over 3–5 minutes. Research shows that 78% of customers do not call back—they immediately buy from a competitor.

2

Skyrocketing Operator Turnover: Average tenure of a first-line call center agent is less than 3 months. Recruiting, onboarding, and workspace expenses cost companies $600 to $1,000+ per month per agent.

3

Inconsistent Quality & Burnout: Human agents suffer from mental fatigue, leading to skipped scripts, inaccurate CRM logging, and slow response times.

4

Multilingual Code-Switching Challenges: In Central Asia and many international hubs, customers frequently blend languages mid-sentence (e.g. Kazakh and Russian, or Spanish and English). Rigid IVR keypads («press 1 for English») cause frustration and high drop-off rates.

5

Lack of In-Call Actions: Traditional agents take several minutes to generate an invoice or payment link. In 2026, buyers expect instant 1-click checkout (via Kaspi QR in Kazakhstan or Stripe / Apple Pay globally) delivered directly during the conversation.

According to Alexander Murdasov (Founder of the VibeCode World developer community and MurdaSoft studio, ex-PM at Kcell B2B, NeuroKids.kz ~$6M valuation), implementing autonomous real-time Voice AI agents slashes call handling costs by 80%, provides instant 1-second response 24/7, and boosts conversion-to-payment by 30–45%.

```

┌───────────────────────────────┐

│ Inbound Customer Call │

│ (SIP Trunk / Twilio / Kcell / │

│ Kazakhtelecom / Zadarma) │

└───────────────┬───────────────┘

│

▼

┌───────────────────────────────┐

│ Asterisk / FreeSWITCH / WebRTC│

│ Low-Latency PCM16k Audio Stream│

└───────────────┬───────────────┘

│

▼

┌───────────────────────────────┐

│ Silero VAD (Voice Detection) │

│ Real-time Barge-in Handling │

└───────────────┬───────────────┘

│

▼

┌───────────────────────────────┐

│ Streaming STT (Whisper v3 / │

│ Deepgram with Code-Switching) │

└───────────────┬───────────────┘

│

▼

┌───────────────────────────────┐

│ Voice Agent Orchestrator │

│ (LangGraph / State Machine) │

└───────┬───────────────┬───────┘

│ │

┌────────────────────────┘ └───────────────────────┐

▼ ▼

[Function Calling & Tools] [Ultra-Low Latency TTS]

• ERP / PostgreSQL (Inventory & Price) Cartesia Sonic / ElevenLabs

• Kaspi Pay / Stripe API (Instant Invoicing) Streaming Audio Bytes

• WhatsApp / Twilio SMS (Payment Link Push) End-to-End Latency < 350 ms

• CRM (HubSpot, AmoCRM, Bitrix24, Salesforce) │

│ │

└────────────────────────┬───────────────────────────────────────┘

│

▼

┌───────────────────────────────┐

│ Real-time Voice Response │

│ Streamed Back to Caller │

└───────────────────────────────┘

```


Next-Generation Voice AI Architecture: Sub-450ms Latency

The defining factor of conversational AI is latency. If the system pauses longer than 800 ms, the conversation feels robotic, leading to conversational collisions and dropped calls. To achieve sub-450ms human-like conversational responsiveness, MurdaSoft implements a high-performance stack:

1

Neural Voice Activity Detection (VAD) & Instant Barge-in:

Using Silero VAD, the agent detects user speech interruptions within 50 ms. If the customer interrupts («Wait, do you have that in stock in Almaty?»), audio playback is halted immediately, audio buffers are flushed, and the orchestrator pivots context naturally.

2

Streaming Speech-to-Text (STT):

Audio frames are streamed via WebSockets in 80–120ms chunks to fine-tuned Whisper v3 or Deepgram Nova-2 models. The acoustic model is specially tuned for local dialects, brand names, addresses, and seamless Kazakh-Russian code-switching.

3

Deterministic LangGraph Orchestrator:

The conversational agent operates within a predictable state machine: Greeting → Qualification → Inventory Check → Instant Billing → Confirmation. Strict anti-hallucination guardrails ground all answers in real-time ERP/database data.

4

Ultra-Low Latency Streaming TTS:

Modern neural TTS engines (Cartesia Sonic, ElevenLabs Turbo v2.5) begin streaming synthesized audio chunks back through the SIP/RTP bridge within 150 ms of the LLM generating its first token, delivering lifelike inflection, breathing, and pitch.


In-Call End-to-End Workflow: Call → Qualification → Payment → CRM

Here is how an automated high-intent call operates in practice:

  • •Caller: "Hi! I'm looking for a 5kW backup generator for our facility. Do you have units in your Almaty warehouse and how much does it cost?"
  • •Voice Agent: "Hello! Yes, we have 5kW gasoline generators in stock at our Almaty warehouse, priced at 185,000 KZT ($380 USD). Would you like me to schedule same-day courier delivery?"
  • •Caller: "Yes, please. Can I pay right now via Kaspi QR or card?"
  • •Voice Agent: "Absolutely. I have just issued an electronic invoice to your phone number via Kaspi Pay, and also forwarded the direct payment link to your WhatsApp. You can approve it in 1 tap right now. As soon as payment clears, our fulfillment team will pack your order."

During this 20-second exchange, the agent executed 4 asynchronous tools without human intervention:

1

check_warehouse_stock("generator_5kw", city="almaty")

2

create_crm_deal("Generators", value=185000, stage="invoice_sent")

3

kaspi_pay_issue_remote_invoice(phone="+7701XXXXXXX", amount=185000)

4

send_whatsapp_message(phone, text="Here is your payment link...")


Implementation Pricing, Timelines & ROI (2026)

TierTech StackDelivery TimelinePrice (KZT ₸)Price (USD $)Business ROI
Starter Voice Agent (Inbound Leads / Booking)Asterisk / Zadarma, Whisper STT, ElevenLabs TTS, CRM Integration7–10 business daysfrom 450,000 ₸from $950100% call capture rate 24/7, zero missed opportunities
Bilingual Commerce AI (KK/RU Code-Switching + Kaspi QR)SIP Trunk, fine-tuned Whisper, LangGraph, Kaspi Pay API, WhatsApp Business API14–20 business daysfrom 950,000 ₸from $2,000In-call checkout, 35–45% increase in lead-to-payment conversion
Enterprise AI Call Center (1C ERP, Omnichannel Sync)FreeSWITCH cluster, 50,000+ SKU RAG catalog, 1C ERP sync, AmoCRM / Bitrix243–5 weeksfrom 1,800,000 ₸from $3,80070% reduction in call center payroll, handles 100+ concurrent lines
Global SaaS Voice Bot (EN/ES/DE, Twilio, Stripe)WebRTC streaming, Twilio Voice, Cartesia Sonic, Stripe Billing, HubSpot/Salesforce2–4 weeksfrom $3,000from $3,000Immediate international expansion without hiring foreign reps
Dedicated On-Premise Voice ClusterLocal Kazakhstan Data Center (Law № 94-V), On-Premise LLM (Qwen/Llama 3), Whisper Local4–6 weeksfrom 3,200,000 ₸from $6,800Full regulatory data compliance, zero external API leakage

Step-by-Step 7-Stage Implementation Roadmap

1

Telephony & Workflow Audit (Days 1–2): Review existing telephony infrastructure (Kazakhtelecom, Kcell, Beeline, Zadarma, Twilio, or cloud PBX). Define high-priority conversational paths.

2

SIP / WebRTC Gateway Deployment (Days 2–3): Deploy hardened Asterisk / FreeSWITCH instances with optimized RTP buffers to eliminate jitter and packet loss.

3

Speech Model Calibration & Voice Tuning (Days 3–4): Fine-tune Whisper STT for localized brand vocabulary, Central Asian topography, and bilingual code-switching. Clone or select branded neural TTS voices.

4

LangGraph State Machine & RAG Integration (Days 3–5): Build deterministic dialogue flows with anti-hallucination boundaries locked to verified company data.

5

Fintech & CRM Integration (Days 3–4): Connect payment APIs (Kaspi Pay, Freedom Pay, Stripe) and bidirectional synchronization with CRM systems (AmoCRM, Bitrix24, HubSpot).

6

Human-in-the-Loop Pilot Testing (Days 3–5): Run 300–500 live test calls. Evaluate intent recognition accuracy and fine-tune VAD barge-in sensitivity.

7

Production Launch & Real-Time Telemetry (Day 1): Transition live call queues to the autonomous Voice AI cluster with automated uptime and conversion dashboards.


Frequently Asked Questions (FAQ)

Can the Voice AI handle code-switching between Kazakh and Russian?

Yes. Unlike generic off-the-shelf assistants, our engine incorporates specialized normalization layers and fine-tuned Whisper acoustic weights trained on real-world bilingual dialogues in Kazakhstan. It seamlessly understands mixed sentences, local slang, and street names across Almaty, Astana, and Shymkent, responding in the customer's preferred language.

How does the customer complete payment during the call?

While on the phone, the agent triggers a secure API call to Kaspi Pay (or Stripe for international customers). The caller receives an instant push notification in their banking app or a deep link via WhatsApp/SMS, enabling one-tap payment confirmation before the call finishes.

How does modern Voice AI differ from traditional interactive voice response (IVR)?

Legacy IVR forces users through cumbersome keypad trees («press 1 for sales») and cannot process natural speech. Modern Voice AI conducts natural, bidirectional conversations, handles interruptions gracefully, understands complex queries, and executes real backend actions like booking slots and processing payments.

Is the Voice AI compliant with Kazakhstan's Personal Data Law (No. 94-V)?

Yes. We provide on-premise and local cloud deployment options hosted strictly within certified data centers in Kazakhstan (Almaty and Astana), ensuring full compliance with national personal data sovereignty legislation.

How can I get started with MurdaSoft Voice AI?

Explore our interactive portfolio, architecture patterns, and technical capabilities on the MurdaSoft Vibe Map. To discuss your call center requirements or schedule a live architecture demo, contact founder Alexander Murdasov on Telegram at @murdasoft or @Allexndr.

Need a website or Telegram bot?

Free consultation - let's discuss your project.

Discuss project→