Case Study: Custom Voice AI Agent for a Singaporean Telco
A real-time AI voice agent that converses in Hokkien, Teochew, Cantonese and Singlish — handling plan enquiries, bill information and proactive outreach for a Singapore telco's dialect-speaking customers.
The Problem
Dialect-speaking seniors were underserved by existing call centre systems, creating language barriers and high volumes of repetitive, manually-handled queries.
Language Barriers for Seniors
Elderly customers who speak only Hokkien, Teochew, Cantonese or Hakka could not be served effectively by agents who lacked dialect proficiency.
High Volume of Repetitive Calls
Standard enquiries around billing, plan recommendations, port-in and data roaming consumed large amounts of frontline staff time.
Manpower Constraints
The telco was limited in how many multilingual agents it could deploy, capping its ability to scale support across diverse language groups.
No Dialect-Capable Automation
Existing IVR and chatbot systems only supported English and Mandarin, completely missing senior dialect-speaking customers.
Code-Switching Not Handled
Customers frequently mixed languages mid-sentence — e.g. Cantonese–Mandarin or Singlish–Hokkien — which generic tools could not process.
Missed Proactive Outreach
Payment reminders and promotions could not be delivered in customers' preferred dialect, reducing engagement among senior demographics.
Our Solution
Analytico Voice was customised end-to-end — from knowledge base dialect transformation to CRM integration — to serve the telco's full customer base in their own language.
50+ Languages & Dialects
The voice agent understands and responds in Hokkien, Teochew, Cantonese, Singlish, Mandarin, Tamil and Bengali, covering the full spectrum of the telco's customer base.
Natural Code-Switching
Handles mixed-language inputs such as Cantonese–Mandarin and Singlish–Hokkien, matching the way customers naturally speak rather than forcing a single language.
RAG-Powered Knowledge Base
The telco's FAQ and domain knowledge was cleaned, transformed into dialect-friendly sentence structures and embedded into a vector database for accurate, grounded responses.
Multi-Channel Deployment
Deployed across native phone calls (Twilio), the website voice chat widget and WhatsApp voice messages — meeting customers on their preferred platform.
Real-Time STT, NLU & TTS
Qwen and Gemini APIs power end-to-end real-time speech processing — speech-to-text, intent detection and expressive text-to-speech — with responses in under 3 seconds.
Tone & Sentiment Sensitivity
Recognises politeness markers and indirect speech common in Singapore dialects — e.g. "Can waive the late fee or not?" — and adjusts response tone to match urgency and sentiment.
CRM & Automation Integration
A custom API connects the voice agent to the telco's customer management systems, enabling automated bill reminders, renewal nudges and promotions delivered in the customer's dialect.
Safety & Data Compliance
End-to-end encryption, PDPA compliance, no PII used for model training, data stored within Singapore, and adversarial testing via IMDA Project Moonshot frameworks.
Real-Time Admin Dashboard
Visualises anonymised call logs, query topics by dialect, sentiment trends and non-answer rates — enabling continuous optimisation of the knowledge base and models.
Human Agent Handover
For three consecutive complex queries with low relevance scores, customers are offered seamless escalation to a human agent with full conversation context preserved.
Results
95% Dialect Accuracy
Tested with senior community members, the voice agent achieved 95% accuracy in audio understanding and response quality for Hokkien, Teochew and Cantonese.
3-Second Response Time
All voice responses are generated and delivered within 3 seconds, enabling fluid, natural back-and-forth dialogue without noticeable lag.
Estimated 17–20% Cost Savings
Modelled against current staffing and rental costs, the voice agent is projected to reduce customer service overhead by 17% at 1,000 users and 20% at 10,000 users.
Positive Dialect Speaker Feedback
"This is probably the best level a non-human can achieve in terms of dialect." — Cantonese speaker. Hokkien and Teochew testers also confirmed natural, accurate responses.