Large-model understanding
Understands free-form answers, accents and digressions, and asks a clarifying question when unsure — no rigid keyword trees.
Financial AI · Shared calling platform
LLM Voice Agent
Every outbound call,
a conversation that closes the loop.
One large-model calling platform shared by collections, telesales and customer service. It understands what customers mean, answers with approved scripts, handles interruptions and emotions in real time — and writes every outcome back to your systems.
- CollectionsReminders, promises to pay, hardship
- TelesalesLimit increases, reactivation, cross-sell
- ServiceNotices, follow-ups, surveys
- Business lines on one platform
- 0
- Typical response time per turn
- <0s
- Concurrent calls per cluster
- 0K+
- Calls recorded and inspected
- 0%
How it works
Six building blocks. One call loop.
From the moment your collections system hands over a list to the moment the result comes back, every step runs on one stack — engagement hub, voice control, telephony, gateway, speech engines and the large-model dialogue robot.
One call, from list to result
The upstream system pushes a list; the hub prepares it and hands it to voice control; telephony dials out. Once the customer answers, speech flows through ASR to the dialogue robot and back through TTS — turn after turn until hang-up.
Product architecture
Six modules, each independently deployable and scalable. Select a module to see what it does.
Customer phone
PSTN / mobile
SIP trunk · RTP
SIP / RTP media
ASR text →← reply + action
List push · results back
WebSocket streaming audio
How the dialogue robot decides what to say
The large model understands; your approved knowledge base speaks. Each turn is routed to the right robot, matched to an intent, then mapped to a vetted script and action — so answers are natural, but never improvised on regulated topics.
VOICE CONTROL
ASRCustomer speech → text
TTSTemplate audio + parameter TTS
- 1RobotId routingQuery in · pick the robot for this call Collections robotLimit-increase robotService robot
- 2Load robot configLocate prompt template, knowledge base and model
- 3Build promptPersona, task, dialogue context and candidate intents
- 4LLMIntent recognition · emotion · barge-in semantics
- 5Intent tage.g. promise_to_pay · ask_rate · request_agent
- 6Script & action outputApproved reply + action: continue, transfer, SMS, hang up
ROBOTID CONTEXT
- Prompt
- Model endpoint
- Knowledge base ID
- Dialogue context
KNOWLEDGE BASE
- Q&A scripts
- Script actions
- Parameters & variables
- Recording files
Unified vector store
VOICE ADMIN
Robot configurationOptional — defaults apply if unset
Knowledge baseRequired for every robot
Recording assets
Variables
PRIVATE MODEL
Base model+
Fine-tuned weights→
Model instance on your GPUs
Run Qwen, DeepSeek or GLM privately, fine-tuned on your own call transcripts — or route to any model through the naviai AI Gateway.
Business capability layers
Business teams work in the application layer; voice control and the engines underneath are shared by every line of business.
APPLICATION
Agent configuration
Prompt instructionsDialogue simulationASR/TTS engineTransfer rulesSpecial eventsKnowledge & tag libraries
Calling tasks
Call listsReports & BIAgent selectionLine selectionDialling strategyCall records
Knowledge centre
Q&A knowledgeDocument knowledgeTop questions
VOICE CONTROL
Speech recognitionBarge-inTurn & duration controlFallback handlingVoice-assistant detectionSpeech synthesisSilence detectionEmotion detectionNoise filteringSIP proxy
TELEPHONY PAAS
FreeSWITCH clusterSIP proxyCTIVOSCDR & recordingIVR
AI ENGINE · DM & NLU
Knowledge graph & FAQDialogue logicIntent understanding & clarificationRobot task flows
Large models + core AI technology
LABEL & TRAIN
ASR / TTS labellingIntent labellingQ&A training
RESOURCE ADAPTERS
ASR adapterTTS adapterASR / TTS enginesModel adapterQwen · DeepSeek · GLM
Technical architecture
Stateless services on Kubernetes, clustered storage and shared middleware — built to scale out for peak collection days.
Business
Collections systemTelesales systemCustomer service system
Outbound hub
Configuration
RobotsRobot rulesRecording assetsVariables
Calling APIs
Task submissionResult callbackStop-calling
Call records
BatchesTask detailsCall details
Voice link
VCM system
Business management service
BatchesReal-time tasksHistoryAnti-harassmentBackground audio
Real-time call service
External API service
System configuration service
Speech platform
ASRTTS
Large model
Intent recognitionKnowledge base
Call centre
Outbound callsLive transferCall records
Storage
MySQL cluster
Distributed file storage
Elasticsearch cluster
Redis cluster
Shared components
Message push File service Config centre Message queue Job schedulerSee a call unfold
One platform. Three very different conversations.
Pick a business line and watch what the agent hears, understands and does — intent, emotion, latency and actions in real time.
call://collections
00:00
Core capabilities
Sounds human. Behaves like your best-trained agent.
Feature-complete against the leading AI calling platforms in Asia and beyond — with the controls a regulated lender needs.
Compliant by design
The model picks the intent; replies come from approved templates with amounts and dates filled by parameter TTS. Nothing regulated is improvised.
Sub-second turn-taking
Streaming ASR, early intent detection, cached template audio and first-packet TTS keep each reply under a second.
Barge-in, silence & noise
Stops speaking the moment a customer interrupts, re-prompts after silence, filters background noise and detects voice assistants and voicemail.
Emotion-aware handoff
Detects frustration, distress or complaints and transfers to a live agent mid-call — with a summary, so the customer never repeats themselves.
Calling strategy engine
Calling windows, pacing, retry rules, frequency caps, blacklists and do-not-call checks, line selection and caller-ID rotation — per campaign.
Multilingual across APAC
English, Mandarin, Cantonese, Malay, Indonesian, Thai and Vietnamese voices and ASR, with local hot-word libraries.
Recording, QA & reports
Every call recorded, transcribed and quality-inspected; outcomes, PTP rates and costs flow to dashboards in XinBI.
Train, test, improve
Simulate dialogues before launch, label ASR and intents from real calls, and retrain Q&A — accuracy goes up every week.
Built once, shared by every line
A shared platform for the whole customer lifecycle.
Each business line gets its own robots, knowledge base and strategy — on shared lines, engines and governance. A new scenario goes live in days, not quarters.
Collections
- Pre-due and early-stage reminders
- Promise-to-pay capture with SMS confirmation
- Hardship detection and partial-payment offers
- Stop-calling the moment a payment lands
+18%PTP rate vs. IVR
−60%cost per contact
Telesales
- Credit-limit increase and loan top-up offers
- Dormant-customer reactivation
- Lead qualification and warm transfer
- Follow-up via WhatsApp or SMS
3×leads screened per day
+25%conversion after handoff
Customer service
- Repayment, KYC and account notices
- Complaint follow-up and return visits
- Satisfaction surveys
- Fraud and suspicious-activity verification
24/7within permitted hours
90%+calls completed without an agent
Figures are typical results from comparable deployments and depend on your portfolio, lines and scripts.
Why naviai
Beyond the IVR robot.
Rule-based voice robotnaviai LLM Voice Agent
UnderstandingKeywords and fixed flow treesLarge-model intent understanding with clarification
New scenarioWeeks of flow configurationPrompt + knowledge base, live in days
AnswersRigid, one path per nodeNatural, yet always from approved, parameterised templates
Interruptions & emotionTalks over the customerBarge-in, silence handling, emotion-aware transfer
StrategySeparate dialler, manual listsBuilt-in windows, pacing, retries, DNC and line routing
ComplianceSampled recordings100% recording, transcription, QA and audit trail
DeploymentVendor cloud onlyCloud or fully on-premise, with private models and your own lines
Frequently asked questions
QCan the robot say something non-compliant?
No. The large model decides what the customer means; what the robot says comes from scripts your compliance team approved, with amounts, dates and names filled in as parameters. Sensitive topics can be routed straight to a human.
QHow does it connect to our collections or CRM system?
Through three APIs: task submission to push lists, result callbacks for outcomes and recordings, and stop-calling to cancel pending calls instantly — for example when a payment arrives.
QCan we use our own phone lines and numbers?
Yes. The voice gateway connects any SIP trunk or carrier, routes each call to the best line by region, cost and answer rate, and handles rates and billing per line.
QHow do you keep calling within regulations?
The strategy engine enforces permitted calling hours, frequency caps, consent and do-not-call lists per market, and every call is recorded and inspected so you can prove it.
QWhere do recordings and transcripts live?
Wherever you need them: in our Singapore cloud, a dedicated cloud, or entirely on-premise with private ASR, TTS and large models — no audio has to leave your network.
Hear it on your own scripts.
Send us one script from collections, telesales or service. We'll build the robot and call you with it — in your language, on your scenario.
What a pilot looks like
- Week 1 · scripts, knowledge base and lines
- Week 2 · simulation testing and compliance review
- Weeks 3–4 · live calls on a pilot segment
- Then · A/B against your current channel