Shunya LabsShunya LabsPlayground

Horizon · Radiance

AI Customer Support Voice Agent

Handle customer conversations with an AI voice agent that can listen, understand, respond, and speak back in real time. Build customer-support voice experiences with Shunya Horizon and Radiance — turning live customer speech into text, passing the conversation to your AI application, and generating natural spoken responses through the same speech stack.

Real-time speech to textStreaming text to speechMultilingual recognitionCode-switch support calls
Customer

Hi, I made a payment yesterday but I still haven't received the confirmation.

AI Agent

I can help you check that. Could you share the transaction or order number?

Listen · Understand · Speak back
Horizon · ListenSupport applicationRadiance · SpeakTurn by turn

AI Customer Support Voice Agent

Customer calls should not have to wait for an Available Agent

Traditional phone support depends on people being available every time a customer calls. Many of theseconversations follow predictable workflows. An AI Customer Support Voice Agent can become theconversational layer between the caller and the systems that contain the information they need.

Customers may contact support to

Requests that follow predictable workflows

  • Check an order
  • Ask about a payment
  • Confirm an appointment
  • Understand a service
  • Update account information
  • Check a request status
  • Ask a product question
  • Get basic troubleshooting help
  • Find the right department
  • Repeat information they have already provided elsewhere

Shunya speech layer

A voice application, not a prerecorded phone menu

The result is not simply a prerecorded phone menu. It is a voice application that can process what the customer actually says and generate a response dynamically.

Customer speaksHorizonSpeech becomes textAI / LLM / Business logicUnderstand the request. Retrieve information. Decide the next response.RadianceResponse becomes speechCustomer hears the answer
CapabilityCustomer Support Workflow
Real-time Speech to Text
Streaming Text to Speech
Multilingual Speech Recognition
Code-switched Speech Workflows
Speaker-aware Transcription
Word-level Timestamps
Intent & Sentiment Intelligence
Batch Call Transcription
Telephony Audio Formats
API Integration

Try a customer conversation

Test the kind of support conversation yourusers Actually Have

A prepared voice demo cannot tell you how a support workflow will handle your customer terminology, ordernumbers, product names, questions, accents, or conversational style. Test a realistic interaction.

Sample conversation

Customer

Hi, I made a payment yesterday but I still haven't received the confirmation.

AI Agent

I can help you check that. Could you share the transaction or order number?

Customer

Yes, the order number is 48217.

AI Agent

Thank you. Let me check the current status.

Customer voiceSpeech recognitionCustomer requestAI applicationRelevant informationGenerated responseSpeech synthesisCustomer hears the answer

Listen to the customer

Every voice agent starts by understanding WhatWas Said

Before an AI system can answer a customer, the spoken conversation needs to become machine-readable information.Horizon can act as the listening layer. For real-time conversations, speech recognition needs to operate whilethe caller is still interacting with the system. Horizon supports streaming transcription through aWebSocket workflow. Streaming transcription can return interim text while speechis still arriving, followed by finalised transcript segments.

Customer saysI want to know when my order will arrive.HorizonTranscriptI want to know when my order will arrive.

Use streaming STT for

  • Live customer calls
  • AI voice agents
  • Interactive support
  • Agent-assist applications
  • Conversational IVR
  • Real-time service workflows

Listen

The voice agent does not have to wait until the complete call has finished

The transcript can then move into the application layer responsible for understanding the request.

ListenAudio inPartial transcriptFinal transcriptAI application

Understand what the customer wants

The transcript is only the Beginning

Both sentences may contain the word order. But they require completely different actions. The application layer needs to determine whatthe customer is trying to accomplish. The transcript becomes usable conversation data instead of remaining only a block of text.

Order tracking

Where is my order?

Cancellation

I want to cancel my order.

Intent

What is the customer trying to do?

Sentiment

How does the conversation appear?

Emotion

What dominant emotion appears in a speech segment?

Summary

What happened during the conversation?

Speakers

Who spoke when?

Word timestamps

Where did specific words occur in the audio?

Intent examples

  • Track an order
  • Ask about a payment
  • Request cancellation
  • Reschedule an appointment
  • Ask for account information
  • Request support

Sentiment

How does the conversation appear?

PositiveNeutralNegative
SpeechTranscriptStructureIntent, sentiment, emotion, speakers, timing, summaryApplication

Generate the response

The AI decides what should happen Next

Speech recognition answers what the customer said. The application layer determines what should be doneabout it. The exact business-system integrations are controlled by the application buildingthe voice agent. Shunya provides the speech layer around that application.

This layer may contain

  • An LLM
  • Business rules
  • Internal APIs
  • Search
  • Retrieval
  • Customer account systems
  • Order systems
  • Scheduling systems
  • Knowledge bases
  • Other application logic
CustomerWhere is my order?TranscriptWhere is my order?IntentOrder trackingApplicationRetrieve current order statusResponseYour order has been dispatched and is expected tomorrow.

Speak back in real time

A voice agent needs to answer with voice, notanother block of Text

Once the application generates a response, Radiance can convert it into spoken audio. Radiance supportsstreaming synthesis for applications that need speech to begin before the entire audio file has beengenerated. For a live customer-support conversation, streaming is the more relevant workflow.

<350ms

First audio in streaming mode

AI responseYour order has been dispatched and is expected tomorrow.RadianceSpoken responseThe customer hears the generated answer.

Streaming TTS is designed for

  • Voice agents
  • Conversational AI
  • Interactive assistants
  • Customer support
  • Real-time applications

Batch TTS is better suited where the entire audio can be produced before playback

  • Notifications
  • Recorded messages
  • Announcements
  • Pre-generated prompts

One complete customer-support loop

Listen. Understand. Act. Speak.

A customer-support voice agent combines multiple layers. The process can continue turn by turnuntil the workflow is complete. That repeating loop creates the conversational experience.

01 · Customer speaksI need to change my appointment to Friday.02 · HorizonConverts speech into text.03 · AI / ApplicationIdentifies intent: reschedule appointment. Extracts the information required by the workflow.04 · Business systemChecks available appointment options.05 · AI responseI can help with that. I found two available slots on Friday.06 · RadianceGenerates natural spoken audio.07 · Customer hears the response

The loop

ListenHorizonUnderstandAI / LLMActBusiness logicRespondGenerate textSpeakRadianceListen again

More than yes or no

Customers do not speak like IVR Menus

The caller expresses the request naturally. The speech layer converts that conversation into text that theAI application can process. The difference is not simply voice quality. It is the ability to processan open spoken request instead of forcing every conversation into a fixed menu tree.

Press 1

Press 1 for orders.

Press 2

Press 2 for payments.

Press 3

Press 3 for appointments.

I paid yesterday but the amount is still showing as pending.

My order was supposed to arrive today. Can you tell me where it is?

I need to move tomorrow's appointment to next week.

Traditional menu

  • Customer
  • Fixed choices
  • Keypad selection
  • Another menu

Conversational workflow

  • Customer
  • Natural speech
  • Speech recognition
  • AI understanding
  • Relevant workflow

Multilingual customer support

Your customers may not all speak the Same Language

Customer-support operations often serve users across regions, countries, and language preferences. Horizon isdesigned around multilingual speech recognition. The Shunya speech platform supports broad multilingualtranscription and specialised models for Indian-language and code-switched speech. A supportapplication can therefore choose a speech-recognition path according to the conversation.

Global-language speech

Use the appropriate multilingual recognition workflow.

Indian-language speech

Use speech recognition designed for supported Indian languages and speech patterns.

Mixed-language speech

Use a code-switching workflow where the conversation naturally moves between languages.

ExampleI made the payment yesterday लेकिन confirmation अभी तक नहीं आया. The customer is not speaking two separate conversations. It is one continuous request. The speech-recognition layer should preserve that conversational context.

Hinglish and code-switched support

Real customer conversations do not always stay insideOne Language

In India, support conversations frequently combine English with another language. Horizon Code-Switch is designedfor conversations that naturally move between languages. A customer should not have to change the waythey naturally speak simply because the support system expects one language.

Hindi + English

मेरा payment successful दिखा रहा है but order confirm नहीं हुआ.

Tamil + English

என் order dispatch ஆயிடுச்சா?

Telugu + English

నా payment complete అయింది but confirmation రాలేదు.

Users may mix

  • Product names
  • English technical terminology
  • Payment vocabulary
  • Brand names
  • Order IDs
  • Addresses
  • Native-language conversation
Mixed-language speechHorizon Code-SwitchContinuous transcriptAI application

Speaker-aware customer calls

Know which side of the conversation Said What

When a call contains more than one participant, a plain transcript can lose important structure.Where configured, speaker identification can also resolve registered voices against knownspeaker profiles. The result is a conversation that retains more of its original structure.

Without speaker separation

Hi, how can I help you today? My payment is still showing as pending. Let me check that for you.

With speaker diarization

SPEAKER_00 Hi, how can I help you today?
SPEAKER_01 My payment is still showing as pending.
SPEAKER_00 Let me check that for you.

Speaker diarization can make support transcripts more useful for

  • Call review
  • Quality monitoring
  • Conversation analytics
  • Training
  • Search
  • Summaries
  • Escalation review
  • Post-call workflows

Know when something was said

Connect the transcript back to the Call

Horizon can return word-level timing information. The transcript becomes a map back to the conversation.

WordStartEndConfidence
payment4.21s4.68s0.98
pending5.03s5.51s0.97
yesterday6.11s6.72s0.96
Search“payment pending”FindMatching transcript segmentJumpRelevant moment in the audio

Word timestamps can support

  • Call playback
  • Transcript highlighting
  • Search
  • QA review
  • Dispute investigation
  • Training data
  • Conversation analytics

After the call

The conversation can remain useful after thecustomer Hangs Up

Real-time support is one part of the workflow. Recorded calls can also be processed throughbatch transcription. This allows the same speech platform to support bothduring the conversation and after the conversation.

A completed conversation can become

  • A searchable transcript
  • A speaker-labelled record
  • A summary
  • An intent classification
  • Sentiment information
  • Structured conversation data
  • Input for another application

During the conversation

Real-time voice interaction

After the conversation, the same stack supports analysis and operational workflows.

After the call

Call endsRecorded audioHorizon batch workflowTranscriptSpeakers + timestampsOptional intelligenceCRM / Search / Analytics / Review

Customer support use cases

One voice agent can support many differentCustomer Journeys

Order Tracking

Understand questions such as “Where is my order?” Retrieve the relevant status through the application workflow and speak the answer back.

Appointment Support

Handle conversational requests involving booking, confirmation, rescheduling, availability, and reminders.

Payment Support

Understand questions about payment confirmation, pending transactions, payment status, failed transactions, and general billing information. The application determines which account or transaction systems are queried.

Product Questions

Answer questions using the business's approved product or service information.

Account Support

Guide users through supported account-related workflows where the application has access to the required systems.

Service Requests

Capture what the customer needs and route the request into the relevant workflow.

Frequently Asked Questions

Handle repeated questions without forcing customers to navigate several IVR levels.

After-Hours Support

Create voice experiences that are not dependent on a human agent being available for every supported request.

Call Routing

Understand the reason for the call before routing it to the appropriate workflow or team.

Human handoff

AI does not need to handle every Conversation

A useful customer-support system needs to know when another workflow should take over. The applicationcan define rules for escalation. The speech layer and the business decision layer should remainseparate. Shunya handles speech recognition and speech generation. Your applicationdecides what actions, routing rules, and escalation logic should apply.

Examples may include

  • The requested workflow is not supported
  • The customer asks for a human
  • Required information cannot be resolved
  • Business rules require manual review
  • The conversation needs a specialised team
  • The application reaches a defined escalation condition

Escalation workflow

Voice agentHandles supported workflowEscalation conditionDefined by your applicationHuman / specialist workflow

Telephony-ready audio

Customer-support voice agents eventually need towork with Phone Audio

Web audio and telephone audio do not always use the same formats. Horizon and Radiance support audio formats suited to real-time andtelephony workflows. Real-time transcription can work with representations suited to live speech and telephony. This allowsthe speech layer to sit inside voice applications that interact with telephone audio rather than only browser recordings.

Radiance output formatsCommon use
PCMReal-time and low-latency audio
WAVHigh-quality audio workflows
MP3Web playback and general applications
OGG OpusEfficient streaming and web audio
FLACLossless audio
μ-lawCommonly used in telephone and SIP environments
A-lawUsed by telephony systems based on G.711 A-law

Developers

Build the customer-support conversation as a Pipeline

An AI Customer Support Voice Agent is not one model. It is a sequence of components.

Pipeline

InputLive customer audioSpeech recognitionHorizonTranscriptCustomer requestApplication layerLLM, business logic, retrieval, internal APIs, and customer systemsResponse textGenerated answerSpeech synthesisRadianceOutputSpoken response

Streaming STT

Use a persistent WebSocket connection when transcription needs to happen while the customer is speaking.

Streaming TTS

Generate speech progressively when the customer should begin hearing the response without waiting for the entire audio file.

Batch STT

Process completed calls when the workflow requires post-call transcription or analysis.

Structured output

Depending on the selected STT workflow, the application can work with:

  • Transcript text
  • Segments
  • Language information
  • Speakers
  • Word timestamps
  • Confidence information
  • Enabled conversation-intelligence fields

Choose the speech workflow

The conversation determines the Model

Different support calls may require different speech-recognition paths. Radiance turns the final response into spoken output.

Horizon Universal

For broad multilingual and global-language recognition. Use when the conversation is primarily in a supported global language.

Horizon Indic

For supported Indian languages and Indian speech patterns.

Horizon Code-Switch

For conversations that naturally mix languages.

Use Horizon Indic primarily for

  • Hindi
  • Tamil
  • Telugu
  • Bengali
  • Marathi
  • Kannada
  • Malayalam
  • Gujarati
  • Punjabi
  • Urdu
  • Other supported Indian languages

Code-Switch examples

  • Hinglish
  • Tanglish
  • Telugu + English
  • Bengali + English
  • Marathi + English

The model follows the conversation

Global-language speechUniversal workflowIndian-language speechIndic workflowMixed-language speechCode-Switch workflowAI applicationRadiance

Built for production conversations

A voice demo is not a Customer-Support System

Generating one transcript and one spoken sentence is only the beginning. The speech layer surrounds the intelligence and business logic that make the support experience useful.

Production customer-support applications also need

  • Streaming speech recognition
  • Streaming speech synthesis
  • Multilingual support
  • Code-switching
  • Speaker information
  • Timing information
  • Telephony-compatible audio
  • API integration
  • Post-call transcription
  • Conversation intelligence
  • Application logic
  • Error handling
  • Escalation workflows
RequirementShunya Speech Stack
Real-time STT
Streaming TTS
Batch STT
Multilingual Recognition
Indic Speech Recognition
Code-switched Recognition
Speaker Diarization
Speaker Identification
Word Timestamps
Intent
Sentiment
Emotion
Summarisation
Telephony Audio
HTTP APIs
WebSocket Workflows
Self-hosted STT Deployment

One conversation stack

Listen, understand, act, and speak through the same support conversation.

User speaksHorizonListenAI / LLMUnderstandApplicationActRadianceSpeakUser hears response

Questions

Frequently Asked Questions

What is an AI Customer Support Voice Agent?

An AI Customer Support Voice Agent is a conversational application that can listen to spoken customer requests, convert speech into text, process the request through an AI or application workflow, generate a response, and return that response as speech.

How does an AI voice agent understand a caller?

Speech recognition converts the caller's audio into text. That transcript can then be processed by an LLM, business rules, retrieval system, or another application layer.

Does Shunya provide real-time Speech to Text?

Yes. Horizon supports real-time streaming transcription through WebSocket workflows.

Can the agent respond with speech?

Yes. Radiance can convert generated text into spoken audio and supports streaming synthesis for real-time applications.

Can a customer-support voice agent handle multiple languages?

Shunya's speech platform supports multilingual speech-recognition workflows, including broad global-language coverage and specialised Indian-language models.

Can it understand Hinglish?

Horizon Code-Switch is designed for conversations that naturally move between languages, including mixed Indian-language and English speech.

Can Shunya separate the customer and agent in a call transcript?

Speaker diarization can identify speaker changes and assign labels to transcript segments.

Can I get timestamps for the transcript?

Yes. Verbose Horizon output can include word-level start times, end times, and confidence information.

Can Shunya identify customer intent?

Shunya supports optional conversation-intelligence features including intent, sentiment, emotion, and summarisation in supported transcription workflows.

Can calls be analysed after they finish?

Yes. Recorded calls can be processed through batch transcription for transcript generation and supported post-call intelligence workflows.

Can an AI support agent connect to our order or CRM system?

The business-system integration belongs to the application layer around the Shunya speech stack. Shunya converts speech into usable text and generated responses into audio; your application controls how external or internal systems are queried.

Can the voice agent transfer a customer to a human?

Handoff and escalation logic can be implemented by the application according to the organisation's own rules. The speech layer can provide the conversation input, while the application determines when another workflow should take over.

Does Shunya support telephone audio?

Horizon and Radiance support audio representations used in real-time and telephony workflows, including μ-law and A-law where applicable.

Is the AI Customer Support Voice Agent one Shunya model?

No. A complete voice agent is a pipeline. Horizon provides the listening layer, an AI/application layer handles reasoning and actions, and Radiance provides the speaking layer.

Can the same stack support post-call analytics?

Yes. Completed audio can be processed using batch speech recognition, with supported capabilities such as diarization, timestamps, summaries, intent, sentiment, and other conversation-intelligence fields.

Get Started

Build customer support that can Listen and Speak

Don't evaluate a voice agent using a scripted hello. Test the questions your customers actually ask. Try an order-status question,payment query, appointment change, product question, mixed-language conversation, service request,or support interaction from your real workflow. Use Horizon to listen. Connect your applicationto understand and act. Use Radiance to speak back.