Telemarketing — Vietnam, multi-dialectVoice AI & GraphRAG grounding

Voice automation across Vietnamese dialects, grounded in a knowledge graph

A tonal language, significant regional dialect variation, and product knowledge scattered across PDFs and wikis — automated to 60–70% of outbound calls without letting the agent invent answers.

The path of a single outbound call. Inbound speech across Vietnamese dialect regions plus English enters a model-agnostic speech layer spanning AWS, 11Labs and OpenAI providers, tuned for tone and swappable without re-architecture, alongside an agent core performing intent recognition against a policy grammar. Every candidate response then passes GraphRAG grounding, checked against the knowledge graph so the agent cannot state a product fact the graph does not hold. A confidence threshold splits the path two ways: above it the agent answers and the call continues, which covers 60 to 70 percent of outbound calls; below it the call is handed to a person with full context. There is no third branch where the agent guesses. Consent and do-not-call compliance are enforced in the call flow through the CRM rather than recorded afterwards.
Two outcomes, and neither is a guess. The uncertain call goes to a person with the context already assembled.

Project snapshot

Client
A leading telemarketing and customer service organisation operating across Vietnam’s linguistic regions
Industry
Customer service and outbound telemarketing
Business function
Outbound calling, lead qualification and product consultation
Challenge
Overwhelming volumes of repetitive calls created wait times and agent burnout, service quality varied across dialects, and product knowledge was scattered across PDFs, wikis and spreadsheets. Vietnamese is tonal with significant dialectal difference, so basic translation was never going to be enough.
Solution
A plug-and-play voice kit across AWS, 11Labs and OpenAI STT/TTS providers, tuned for Vietnamese. GraphRAG grounds every response in a knowledge graph so the agent cannot invent product facts. Confidence thresholds trigger human handoff, a dashboard tracks Word Error Rate and latency stratified by accent, and ODOO connectors handle consent and DNC compliance.
Result
60–70% of outbound calls automated with quality held through intelligent escalation, 50–60% reduction in time-to-conversion during proof of concept, and under 2 days to configure a new product or language.

Key outcomes

60–70%
Of outbound calls automated
50–60%
Faster time-to-conversion
< 2 days
To add a product or language

The client

High-volume calling across a linguistically diverse country

A leading telemarketing and customer service organisation running high-volume interactions across Vietnam’s dialect regions and in English.

The brief was intelligent automation that preserved cultural sensitivity and service quality — not automation that flattened both, which is the usual outcome when a voice system meets a language it was not built for.

The challenge

Six problems, and none of them was “build a chatbot”

Volume and burnout. Overwhelming quantities of repetitive calls, producing wait times for customers and burnout for agents.

Regional inconsistency. Customer experience varied wildly across Vietnamese regions and dialects, making consistent delivery close to impossible.

Fragmented knowledge. Product knowledge sat across PDFs, wikis and spreadsheets, so real-time retrieval during a live call was extremely difficult.

No operational visibility into call interactions, conversion patterns or agent performance. No cost-effective way to scale across dialects and English without exponential resource growth. And underneath all of it, tonal and cultural complexity that basic translation does not reach.

Constraints

What we had to design around

  • TonalityVietnamese is tonal. Speech recognition and synthesis both have to respect tone, or the system produces fluent, confident nonsense.
  • DialectPerformance reported as a single national average would hide entire regions failing underneath it, because high-traffic regions dominate the mean.
  • GroundingAn agent that improvises a product fact during a sales call is a compliance problem, not merely a quality one.
  • Vendor riskCommitting to one speech provider in a fast-moving market would have forced a re-architecture within the year.
  • RegulatoryOutbound calling carries consent and Do-Not-Call obligations the system has to enforce, not merely record after the fact.

Our approach

Ground the agent, then make the model swappable

Two decisions did most of the work. Everything the agent says is grounded in a knowledge graph, and no part of the stack is committed to a single model vendor.

The first is a correctness decision: in an outbound sales context, a hallucinated product fact is a regulatory exposure. The second is a longevity decision — speech models were changing fast enough that a hard dependency on one provider would have dated the system within months.

The solution

A GraphRAG-enhanced voice platform

  1. Model-agnostic voice kit — Plug-and-play across AWS, ElevenLabs, Soniox and OpenAI speech providers, tuned for Vietnamese tonal nuance. Models swap without re-architecture.
  2. Conversation grounding — GraphRAG over interconnected knowledge graphs. Every candidate response is checked against the graph, so the agent cannot state a product fact the graph does not hold.
  3. Agent core — A deliberately lightweight LLM system handling intent recognition and policy grammar, with confidence thresholds driving handoff.
  4. Accent-stratified monitoring — Word Error Rate, latency and engagement, reported per dialect region rather than pooled into one number.
  5. Enterprise integration — ODOO CRM connectors for lead updates, call logging, consent management and DNC compliance — inside the call flow.
The voice pipeline agent as deployed. User audio enters a speech-to-text stage backed by AWS, ElevenLabs and Soniox adapters, producing a transcript. The transcript passes to the LLM stage, which returns text to a text-to-speech stage — again AWS, ElevenLabs and OpenAI — producing agent audio back to the caller. Alongside the loop, the application logic layer holds the functions, retrieval-augmented generation and other tools the agent can call.
The deployed pipeline. Each speech stage is an adapter, which is what makes the provider swap a configuration change rather than a rebuild.
The full platform architecture in four bounded regions. Telephony media connects over WebRTC to a VoiceKit orchestrator in the orchestration region. The orchestrator fans out to speech adapters — Amazon Transcribe, Amazon Polly, ElevenLabs, Soniox and OpenAI — and to an AI services region containing LLM intent recognition backed by Amazon Bedrock running Sonnet 4, with session memory in DynamoDB and an S3 archive. A data integration region holds CloudWatch dashboards and a PostgreSQL store, connected to the archive by a retrieval-augmented generation path.
The whole platform. The speech adapters sit in their own bounded region — swapping a vendor never reaches the orchestrator.

Architecturally: Live Kit Voice with LLM intent recognition and tools, running an Understand-Resolve-Confirm interaction framework. Blue-green deployments with shadow testing. RBAC, audit logs and PII masking. Auto-scaling infrastructure with intelligent cost management through model switching.

An illustration of a measurement decision. A single national Word Error Rate is shown as the arrangement we refused to ship, because high-traffic regions dominate the mean and a region can fail invisibly inside it. Below, the stratified dashboard that was shipped: separate rows for Northern, North-central, Central and Southern dialects plus English, each reporting Word Error Rate, latency and engagement score independently. The same instrumentation runs at turn level with session tracing and full replay, alongside blue-green deploys, role-based access control, audit logs and PII masking.
A pooled Word Error Rate would have looked fine while an entire dialect region failed underneath it.
< 2 daysto configure a new product offering or an additional language — the reusable integration pattern is what makes the platform worth having

Responsible by design

Four properties that are structural, not prompted

  • Ground

    The agent cannot state a fact the graph does not hold.

    Enforced structurally through GraphRAG rather than by instructing the model to be careful.

  • Escalate

    Below threshold, a person takes the call.

    There is no third branch where the agent improvises. The uncertain case routes to a human with full context.

  • Stratify

    Nothing in the dashboard is a national average.

    A pooled Word Error Rate lets a dialect region fail invisibly behind higher-traffic regions.

  • Enforce

    Consent and DNC live in the call flow, not the audit log.

    Compliance recorded after the fact is compliance you have already breached.

The escalation branch is the one that decides whether this system is safe. A voice agent that answers everything will answer wrongly some of the time, in a medium where the customer cannot see a confidence score. Confidence thresholds mean the uncertain call becomes a handoff rather than a confident guess.

Results

Measured across the proof of concept

  • 60–70% of outbound telemarketing calls automated, with quality held through confidence-threshold escalation rather than by lowering the bar.
  • 50–60% reduction in time-to-conversion, for lead qualification and product consultation, measured during the proof-of-concept phase.
  • Under 2 days to configure a new product or language, through reusable integration patterns rather than bespoke build.

Operational and governance

Per-accent
analytics confirming consistent performance across Vietnam’s linguistic regions, rather than a national average
Turn-level
metrics with session tracing and full replay, so improvement is driven by evidence not impression

Beyond the numbers

What else changed

The model-agnostic architecture means models can be hot-swapped as the field moves — without re-architecting, and without renegotiating the accuracy baseline every time a provider ships a new version.

Governance came with it: compliance enforcement, data retention policy and proactive security testing, plus intelligent model switching that delivered measurable cost reduction alongside the quality gains rather than trading against them.

Both headline figures are ranges from a proof-of-concept phase, and we report them that way deliberately. The 60–70% automation and 50–60% conversion-time reduction were measured on the pilot corpus, not claimed for the full production call volume.

If your voice automation works in one accent and not the others

It is usually visible immediately once the metrics are stratified — and invisible for as long as they are pooled.