AI · AIC-89

Voice AI Agents for Customer Service

Voice AI Agents for Customer Service is a developer course in building Thai voice agents for customer service, from choosing speech-to-text and text-to-speech and using real-time voice APIs to managing turn-taking, connecting business data and handing calls over to staff. It suits Python developers and contact centre IT teams, and you leave with a tested prototype voice agent.

Updated
From 7,110 THB / person 7,900 −10% excl. VAT 7% · group rates available
PDFDownload the course outline
  • Duration12 hours · 2 days
  • FormatOnsite / live online
  • Next roundOn request
  • CertificateIncluded

Course overview

Many customers still pick up the phone when something is urgent, yet call centres struggle to answer in peak hours and staff repeat the same answers all day. The latest voice AI speaks Thai far more naturally and replies almost instantly, but building a voice agent for real customers is different from a text chatbot. It has to keep latency low, know when the caller has finished speaking, let the caller interrupt, read Thai numbers and names correctly, and know when to hand over to a person.

This course shows learners how to build a Thai voice AI agent for customer service in Python. Learners choose between an STT-LLM-TTS pipeline and a real-time speech-to-speech model, try Thai STT and TTS options, build agents with the Gemini Live API and the OpenAI Realtime API, and tune turn-taking and barge-in. They then connect business data through function calling, design

handover to staff with a case summary, connect phone numbers over SIP, put guardrails and PDPA controls in place, and test with conversations covering many situations. It closes with a capstone voice agent of the learner's own. This is a developer course, unlike AI for Customer Service, which covers AI use at the business level. (2 days, 6 hours per day, 12 hours in total, Intermediate level.)

What you’ll gain

  • Choose a voice agent architecture that suits the service task and its latency budget
  • Select and tune Thai speech-to-text and text-to-speech so they handle numbers and names
  • Build Thai voice agents with the Gemini Live API and the OpenAI Realtime API
  • Tune turn detection and barge-in so conversations feel natural
  • Connect voice agents to business data safely through function calling
  • Design handover to staff with a context summary so customers need not repeat themselves
  • Understand phone connections over SIP and web voice over WebRTC
  • Put guardrails, PDPA controls and a test suite in place before going live with customers

Who this course is for

  • Python developers and AI engineers who need to build voice agents for customer service
  • Call centre and contact centre IT teams who want to adopt voice AI
  • Developers who have built text chatbots and want to extend them to voice
  • Solution architects who must choose a platform and design the telephony integration
  • Technical product owners responsible for voice bot or callbot projects

Prerequisites

  • Working Python skills such as functions, async code and installing packages
  • A basic understanding of using LLMs and writing prompts
  • Basic command-line skills and calling REST APIs
  • A laptop that can run Python and VS Code, plus a headset with a microphone

Curriculum

Course Details

The course runs for 2 days, 6 hours per day (12 hours in total, 09:00-16:00), as lectures with labs built on a single sample order and appointment system throughout. Intermediate level. This is a developer course that builds on AI for Customer Service, which covers AI use in service at the business level, and has learners write Python code for a Thai voice agent covering STT and TTS,

real-time voice APIs, turn-taking, function calling, handover to staff and phone connections. Learners use their own or company API accounts, which are billed by usage (some services offer a free tier or trial). Connecting a real phone number requires a SIP trunk or telephony service that the learner signs up for. Learners take home a lab guide, Python code for every lab, a set of test conversations and a pre-launch checklist.

Day 1 Thai Speech and Your First Voice Agent

Section 1: Voice AI in Customer Service

  • Tasks voice agents handle well: repeat questions, status checks, appointments and triage before transfer
  • Tasks staff should take, such as sensitive complaints and decisions with financial impact
  • An STT-LLM-TTS pipeline compared with a real-time speech-to-speech model
  • The time budget for each step, and what makes a conversation sound unnatural
  • Lab: set up Python and API keys and talk to a sample voice agent

Section 2: Thai Speech-to-Text

  • Thai STT options: cloud speech services, ElevenLabs and the open Typhoon ASR Real-time model
  • Streaming compared with batch, and the effect on latency
  • Common problems: names, product codes, phone numbers and English mixed into Thai
  • Lab: compare transcripts of sample call recordings in several accents

Section 3: Thai Text-to-Speech and Voice Persona

  • Choose TTS by naturalness, speed, price and streaming support
  • Prepare text so Thai numbers, dates, prices and abbreviations are read correctly
  • Voice cloning requires the speaker's consent, and each platform's terms for commercial use
  • Lab: create a set of greeting messages and test several voices

Section 4: Lab: Real-time Voice APIs

  • Gemini Live API: connect over WebSocket and stream audio in and out
  • OpenAI Realtime API: sessions, instructions and choosing a voice
  • Set the Thai language, the scope of answers and the agent's persona
  • Lab: build a Thai voice agent that talks through the browser

Section 5: Turn-taking and Barge-in

  • Voice activity detection and deciding that the caller has finished
  • Barge-in: let the caller interrupt and stop the bot's audio immediately
  • Handle silence, background noise and callers who talk at length
  • Orchestrate the pipeline with open source frameworks such as Pipecat or LiveKit Agents
  • Lab: tune turn detection and test it with live conversations in class

Section 6: Lab: Designing Conversations for Voice

  • Prompts for speech: short sentences, one question at a time and no long lists
  • Verify identity and read details back before taking action
  • Topics the bot must not answer, and how to decline politely
  • Lab: an order status conversation with its exception cases
Day 2 Integration, Handover to Staff and Going Live

Section 7: Lab: Function Calling with Business Data

  • Let the agent call tools to check orders, book appointments and open tickets
  • Answer from manuals and FAQs with RAG, keeping answers short for voice
  • Keep talking while a tool runs so the caller is not left in silence
  • Check permissions before revealing customer data

Section 8: Lab: Handover to Staff

  • Handover triggers: the caller asks for a person, is upset, is out of scope, or the bot is unsure
  • Summarise the context so staff can take over without the caller repeating themselves
  • Transfer to a number or SIP address, or create a ticket or callback when nobody is free
  • Lab: build a handover tool with a conversation summary

Section 9: Phone Numbers and Other Channels

  • Connect a voice agent to a phone number through a SIP trunk or telephony provider
  • The Realtime API over SIP and ready-made platforms such as ElevenLabs Agents
  • Voice on the web and in apps over WebRTC
  • Lab: demonstrate a call into the voice agent through a telephony trial service

Section 10: Guardrails, PDPA and Security

  • Tell callers they are talking to AI and ask for consent before recording
  • PDPA: keep only the audio and transcripts you need, mask personal data and set retention periods
  • Defend against prompt injection and attempts to obtain other customers' data
  • Stop the bot from promising anything that policy does not allow

Section 11: Lab: Testing, Measuring and Improving

  • Test with simulated conversations, including angry callers, background noise and off-topic questions
  • Metrics: latency, self-service resolution rate, handover rate and data accuracy
  • Read transcripts and logs to find where the bot goes wrong
  • Lab: test your own agent and refine its prompts and tools

Section 12: Workshop: Voice Agent Capstone

  • Choose a brief: status enquiries, appointments or triage before transfer
  • Design the conversation, the tools and the handover conditions
  • Build and test end to end
  • Give a live demo with a pilot plan and metrics

Schedule & training options

For individuals — public rounds

No public rounds are open right now. Join the waiting list and we will contact you first when the next round opens, or ask us on LINE. Or call 02-570-8449 or 088-807-9770

For organisations — in-house / private

  • Tailor the content to your team’s tools and projects
  • Your dates, at your office or live online
  • Quotation with tax ID for procurement
Corporate training quote

Instructors

Frequently asked questions

How does a speech-to-text, LLM and text-to-speech pipeline differ from a real-time voice API?

A pipeline converts speech to text, sends it to an LLM for a reply and converts the reply back with text-to-speech. You can choose and tune each part, but latency adds up across the steps. Real-time voice APIs such as the Gemini Live API and OpenAI Realtime API take and return audio over a single connection, so replies are faster and more natural, but you have less control over each step.

How does a voice agent detect that the caller has finished speaking and handle interruptions?

Turn detection relies on voice activity detection of silence, or on a model that judges from meaning whether the speaker has finished. The silence threshold needs tuning: too short and the agent cuts callers off, too long and it feels slow. For interruptions (barge-in), the system must stop playing audio as soon as the caller speaks and drop the unspoken part of its reply from the context so the conversation stays accurate.

How should a voice agent hand a call over to a human agent?

Define clear handover triggers, such as the caller asking for a person, a request outside the agent's scope or repeated low confidence. When handing over, have the agent pass a summary to the staff member with the call, covering who the customer is, what they want and what has already been checked, so the customer does not have to repeat themselves. Tell the caller before transferring them.

Who is the Voice AI Agents for Customer Service course for, and what should I know beforehand?

It suits Python developers and AI engineers building voice agents for customer service, IT teams in call and contact centres, and developers who have built text chatbots and want to move into voice. You should write working Python, including functions and async code, understand how to use LLMs and write prompts, and be comfortable with the command line and REST APIs.

What will I be able to do after the Voice AI Agents for Customer Service course?

You will be able to choose and tune Thai speech-to-text and text-to-speech to handle numbers and names, build voice agents with the Gemini Live API and OpenAI Realtime API, connect business data through function calling, design handover to staff with a context summary, and set up guardrails, PDPA controls and tests before launch. Participants who complete the training receive a Certificate of Completion with its own number, which can be verified online.

What do I need to bring for Voice AI Agents for Customer Service?

Bring a laptop that can run Python and VS Code and a headset with a microphone for voice testing. You use your own or your company's API accounts, charged by usage (some services offer a free tier or trial package). Connecting a real phone number requires a SIP trunk or telephony service that you sign up for yourself.