Voice assistant architecture: build a voice bot on the cloud
A voice bot listens, understands, looks things up and answers aloud. Here is how the pieces fit, which service plays each part on AWS, Azure and Google Cloud, and what to settle before callers use it.
Quick answer
A voice bot chains a few services: a call channel takes the call, speech-to-text turns the caller's words into text, a bot or language model decides the reply and looks up your data, and text-to-speech speaks it back. Every call needs an easy hand-over to a person.
| Best for | High-volume, narrow calls such as order status, appointments and simple support |
|---|---|
| Less suited when | Complex or sensitive calls where a person should talk to the caller from the start |
| Start with | One narrow call flow on a ready-made bot product, tested with real callers' accents |
| Main services on AWS | Amazon Translate, Amazon Connect, Amazon Transcribe, Amazon Lex, Amazon Polly |
| Main services on Azure | Azure Translator, Dynamics 365 Contact Center, Azure Speech, Microsoft Copilot Studio |
| Main services on Google Cloud | Cloud Translation, Contact Center as a Service, Speech-to-Text, Customer Experience Agent Studio, Text-to-Speech |
The idea in plain words
A voice assistant is a chain of small steps. The caller speaks, speech is turned into text, a bot works out what the caller wants and decides what to do, and the reply is turned back into speech. Each step is a service you can swap.
The bot in the middle can be rule-based, with fixed intents and flows, or built on a language model. Rule-based bots are predictable and easy to test. Language-model bots handle open conversation better but need firm limits on what they may say and do.
Calls also need a way in and a way out: a phone number or an in-app audio channel, and an easy hand-over to a person when the bot cannot help.
The architecture
Boxes show the job each part does. Pick a cloud to see the service for each job.
Step by step
- The caller speaksA caller dials a phone number, or speaks into your app. The audio reaches your call service.
- Take the callThe call service answers, plays a greeting, routes the call, records it if you have consent, and can hand the call to a person at any time.
- Speech to textThe audio is turned into text in near real time, so the bot can work with what was said.
- Understand and decideThe bot works out what the caller wants, asks follow-up questions and decides the next step. It can be rule-based or use a language model.
- Look up or actThe bot reads or updates your systems, such as orders, accounts or appointments, through your APIs, and only for tasks you have allowed.
- Text to speechThe reply is turned into spoken audio with a natural voice, in the caller's language, and played back.
- Hand over and improveIf the bot is unsure, or the caller asks, pass the call to a person together with the conversation so far. Review transcripts of failed calls to improve the bot.
Services side by side
The main service for each job on each cloud, with a second option where one exists. Names checked on 23 October 2026.
| Job | AWS | Azure | Google Cloud |
|---|---|---|---|
| Language translation |
|
|
|
| Cloud contact centre |
|
|
|
| Speech to text (transcription) |
|
|
|
| Chatbots and voice bots (intent-based) |
|
|
|
| Text to speech (voice generation) |
|
|
|
Common variations
- Callers speak several languages. Add translation, and test each language end to end. Check which languages and voices each speech service supports.
- A language model handles the conversation. Keep it grounded in approved information (see the RAG pattern), limit what it can do, and add safety checks on what it says.
- Voice inside a mobile or web app, not on the phone. Skip the call service and stream audio from the app to the speech service, then play the reply.
- Calls must be analysed afterwards. Store transcripts and recordings securely and run text analysis for topics, sentiment and quality checks, within your retention rules.
Ready-made bot builder or assemble the parts?
Each cloud offers a bot or contact-centre product that bundles the call channel, speech and bot steps. Or you can connect the speech services and a bot or model yourself.
Ready-made bot or contact-centre product
Good
- Fastest route to a working call flow
- Call handling, hand-over and reporting included
- Less integration work
Watch out
- Less freedom over how the conversation works
- Pricing is usually per minute or per conversation
- You depend on the product's languages and voices
Assemble from speech services
Good
- Pick the best speech or language model for each step
- Full control over conversation logic and data
- Can run inside your own app, not just on phone calls
Watch out
- You build and test the whole chain
- You handle call routing, hand-over and recording
- Delays add up across steps, so speed needs care
Keep in mind
- Speed. Even a pause of a second or two feels long on a call. Test the whole chain end to end, stream speech where possible, keep replies short, and place services in the nearest region.
- Accents and languages. Test with your real callers' accents, background noise and mixed languages. Support differs by service and language, so measure instead of assuming.
- A way to reach a person. Always offer a hand-over to a human, pass the conversation along, and never trap a caller in a loop.
- Consent and recording. Tell callers when a call is recorded or handled by an automated system, and follow the telecom and privacy rules where your callers live (for example the GDPR, US state recording-consent laws and India's DPDP Act). Keep recordings only as long as you need them.
- Sensitive details. Do not read out or store card numbers, OTPs or passwords. Mask them in transcripts and logs.
- Phone numbers and calling in India. Check whether your chosen call service supports the phone numbers and calling you need in India, and which telecom rules apply.
- Cost and quality. Phone minutes, speech minutes and any model usage add up for every call. Measure the cost per call on a pilot, and track how many calls the bot solves without a person, together with caller satisfaction.
Words used on this page
Services in this pattern that have changed
Cloud providers rename and retire services often. These affect the parts above. Checked on 23 October 2026.
Sources and further reading
Check details such as features, limits, prices and regions in the providers' own documentation.
- AWS documentation (opens in a new tab)
- Microsoft Azure documentation (opens in a new tab)
- Google Cloud documentation (opens in a new tab)
Announcements about the changed services above
Frequently asked questions
What is a voice bot?
A voice bot is software that talks with people by voice. It turns speech into text, decides what to do, and answers with synthetic speech, often on phone calls or inside an app.
Do I need a contact-centre service?
Only for phone calls. A contact-centre or call service gives you phone numbers, call routing and hand-over to people. For voice inside your own app you can connect the speech and bot services directly.
How well does speech recognition handle Indian languages and accents?
Support varies by service and by language, and it keeps changing. Test each candidate with recordings of your own callers before you choose.
Can a voice bot use a language model?
Yes. A model can handle more open conversation, but it can also say wrong things. Ground it in approved information, limit what it can do, and add safety checks.
How much does a voice bot cost?
The cost per call combines phone minutes, speech services and the bot or model. Start with one type of call, track the cost per call and the share handled without a person, and compare that with each provider's pricing calculator.
Want this reviewed for your case?
CloudBrum architects can adapt this pattern to your documents, users, rules and budget, and estimate the cost. Nothing is sent until you choose.
Building something else?
Describe it in plain words and get matching services on all three clouds.
Service suggester → Cloud consulting →