Shunya LabsShunya LabsPlayground

Horizon · Radiance

Multilingual AI Voice Agent

Build AI voice agents that can listen, understand, and respond across multiple languages without rebuilding the entire voice stack for every market.

Use Shunya Horizon and Radiance as the speech layer for multilingual conversational AI - with global speech recognition, Indian-language models, code-switched conversations, automatic language detection, streaming transcription, and multilingual speech synthesis.

Horizon global + IndicCode-Switch mixed speechAuto-detect language<350ms first audio
User · English

Can you check the status of my order?

User · Hindi

मेरा order कब deliver होगा?

Same application · different speech path
EnglishHindiHinglishStreaming

Multilingual AI Voice Agent

One voice architecture for conversationsacross Languages

A global voice application may need to handle customers who speak:

Languages

  • English
  • Hindi
  • Tamil
  • Telugu
  • Bengali
  • Marathi
  • Spanish
  • French
  • Japanese
  • Korean
  • Other supported languages

The challenge is not only translating a user interface. The application needs to listen to real speech.

That speech may include

  • Regional accents
  • Mixed-language vocabulary
  • Names
  • Product terminology
  • Numbers
  • Telephone audio
  • Interruptions
  • Language switching

A Multilingual AI Voice Agent can use one conversational architecture while selecting the speech-recognition and speech-synthesis path that fits the actual conversation.

User speaksHorizonRecognise supported languageAI / LLM / applicationUnderstand request. Apply business logic. Generate response.RadianceGenerate spoken outputUser hears response
Multiple languagesHorizonAI applicationRadianceMultiple language experiences
CapabilityShunya speech stack
Global Speech Recognition
Indian-language Recognition
Automatic Language Detection
Code-switched Recognition
Real-time Streaming STT
Batch STT
Multilingual TTS
Streaming TTS
Batch TTS
Cross-lingual Synthesis
Mixed-language TTS
Speaker Diarization
Word-level Timestamps
Telephony Audio
API Integration

Try a multilingual conversation

Test the languages your usersActually Speak

A multilingual voice agent should not be judged only on one prepared English sentence. Test realistic customer interactions.

User · English

Can you check the status of my order?

User · Hindi

मेरा order कब deliver होगा?

User · Spanish

Quiero saber cuándo llegará mi pedido.

User · Hinglish

Mera payment complete hai but confirmation अभी तक नहीं आया.

These conversations may need different speech-recognition paths while still entering the same business application.

User speechLanguage / conversation typeAppropriate Horizon modelAI applicationResponse textAppropriate Radiance workflowUser hears response

Global speech recognition

Understand customers acrossinternational Languages

Horizon Universal is Shunya's model for broad global speech recognition. The supported languages include:

Global languages

  • English
  • Spanish
  • French
  • Japanese
  • Korean
  • Chinese
  • Arabic
  • German
  • Portuguese
  • Russian
  • Turkish
  • Vietnamese
  • Indonesian
  • and other supported languages
Global userSpeaks supported languageHorizon UniversalTranscriptAI application

The same speech-recognition architecture can support multiple international markets rather than requiring an entirely separate ASR stack for each language.

Indian-language voice agents

Build voice applications around howIndia actually Speaks

India requires more than English-only speech recognition. Horizon Indic is designed around Indian languages and Indian speech patterns.

Indian languages

  • Hindi
  • Bengali
  • Tamil
  • Telugu
  • Marathi
  • Kannada
  • Malayalam
  • Gujarati
  • Punjabi
  • Urdu
  • Odia
  • Assamese
  • Nepali
  • Maithili
  • Konkani
  • Kashmiri
  • Dogri
  • Sindhi
  • Bhojpuri
  • and additional supported languages
Indian-language speechHorizon IndicMachine-readable transcriptVoice application

A user should not have to switch to English simply because the software was not designed for regional speech.

Automatic language detection

The application may not knowthe language in Advance

A caller may begin speaking immediately. Horizon supports automatic language detection when a language is not explicitly locked.

User speaksHorizonLanguage detectedTranscriptApplication

This can support experiences where the user does not first select Press 1 for English or Press 2 for Hindi. When the language is known, the application can explicitly provide the language code. Locking a known language can reduce ambiguity, particularly for short audio where less speech context is available.

Auto-detect

Useful when language is unknown.

Language lock

Useful when the application already knows the expected language.

Code-switched conversations

Real multilingual speech does not always stay insideone Language

A customer may start in Hindi and switch to English inside the same sentence.

Hindi + English

मेरा payment complete है but confirmation अभी तक नहीं आया.

Tamil + English

என் order இன்னும் deliver ஆகல.

These are not two separate conversations.

Mixed speechHorizon Code-SwitchContinuous multilingual transcriptAI application

Horizon Code-Switch is designed for conversations that naturally switch languages inside the same sentence or interaction.

Hinglish voice agent

Understand Hindi + Englishas one Conversation

Hinglish is common in:

  • Customer support
  • Payments
  • Sales
  • Appointments
  • E-commerce
  • Banking
  • Everyday voice interactions
Example

Mera order कब deliver होगा?

Example

Payment successful show ho raha hai but confirmation नहीं आया.

Example

Kal ka appointment reschedule karna hai.

Hinglish audioHorizon Code-SwitchContinuous transcriptLLM / applicationRadiance

A voice agent should let the user speak naturally rather than forcing language purity.

Regional language + English

Mixed language goesbeyond Hinglish

Business and digital terminology often remains English even inside regional-language speech.

Tamil + English

என் appointment tomorrow morning confirm ஆகிவிட்டதா?

Telugu + English

నా payment successful అయింది but confirmation రాలేదు.

Bengali + English

আমার order dispatch হয়েছে?

Marathi + English

माझी appointment reschedule करायची आहे.

Regional language + EnglishCode-Switch recognitionStructured transcriptApplication

The user should not need to manually switch recognition models every time the vocabulary changes.

Choose the model that matches the conversation

One multilingual platform does not mean onemodel for Everything

The Horizon family includes different recognition paths.

Horizon Universal

For broad international-language speech. Best suited to supported global-language conversations.

Horizon Indic

For Indian-language speech and regional conversational patterns.

Horizon Code-Switch

For conversations where supported languages naturally mix.

Horizon Medical

For specialised healthcare speech where the supported medical workflow is appropriate.

Global languageUniversalIndian languageIndicMixed languageCode-SwitchClinical speechMed

The application should choose the model according to the actual audio instead of treating multilingual speech as one generic category.

One application, multiple languages

Keep business logic separate from languageHandling

The same customer-support workflow may need to work for several customers.

Customer A

English

Customer B

Hindi

Customer C

Spanish

Customer D

Tamil-English

The business task may be identical: check order status.

Speech languageChangesHorizonAdapts recognition pathApplication logicRemains the sameRadianceReturns appropriate speech

This separation allows language expansion without rebuilding the business workflow from scratch.

Multilingual text to speech

Respond in more thanone Language

Listening is only half of a voice agent. The application must also speak. Radiance is designed as a multilingualspeech platform rather than an English-only voice generator. The supported catalogue coversEnglish, major Indian languages, and extended world-language synthesis.

Response textLanguage selectedRadianceSpoken response

Possible experiences

  • English TTS
  • Hindi TTS
  • Tamil TTS
  • Telugu TTS
  • Bengali TTS
  • Marathi TTS
  • Spanish TTS
  • Japanese TTS
  • and other supported workflows

The exact voices available come from the current deployment, not from a generic language list.

Cross-lingual synthesis

One voice platform acrosssupported Languages

Radiance documents cross-lingual synthesis. This allows supported voice workflowsto extend across languages without replacing the speech platform.

ApplicationGenerates responseLanguageEnglish / Hindi / other supported languageRadianceSpoken output

The exact voice behaviour and available speaker identities depend on the current service configuration. Not every language has its own named voices.

Mixed-language TTS

The generated response can also contain more thanone Language

Multilingual conversations do not only affect recognition. The response may itself mix languages. Radiance documentsmixed-language synthesis for Hindi and English and is positioned for multilingual conversational speech.

Spoken response

आपका appointment tomorrow morning confirm है.

One inputHindi + EnglishRadianceContinuous spoken output

This can produce a more natural conversational experience than splitting every phrase into separate speech engines.

Streaming STT

Understand the user whilethey are still Talking

Voice agents need live transcription. Horizon provides real-time streaming recognition.

Live audioHorizonPartial transcriptFinal transcriptAI application

Streaming is suited to

  • Voice agents
  • Customer support
  • Interactive assistants
  • Contact centers
  • Agent assist
  • Real-time AI applications

The agent does not have to wait until the entire recording is complete.

Streaming TTS

Speak while the AI responseis being Generated

Conversational AI should not wait for an entire audio file before playback begins. Radiance supports streamingsynthesis, with a supported first-audio expectation under 350 ms for relevant workflows.

LLM generates responseRadiance begins synthesisAudio chunks returnUser begins hearing response

Streaming TTS is suited to

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

The result is a conversational pipeline rather than a batch audio-generation workflow.

Complete multilingual voice loop

Listen in one language. Process as text.Speak Back

UserSpeaks supported languageHorizonRecognises speechApplication / LLMUnderstands requestBusiness logicRetrieves or performs relevant actionResponse textGenerated in appropriate languageRadianceSpeech synthesisUserHears response

This basic architecture remains consistent across multilingual experiences.

Language preference

Sometimes the application already knowsthe preferred Language

A customer profile may contain a preferred language. The application can use that context.

Preferred language: Hindi

Configure a Hindi-first speech workflow.

Preferred language: English

The same application can use another speech path.

Customer profileHindiConfigure speech workflowHorizon / RadianceHindi-first experience

The language preference itself belongs to the customer or business system.

Change language during the conversation

A user may ask tocontinue Differently

A multilingual application can use that request to change the configured speech path where supported.

Customer

Can we continue in Hindi?

User requestHorizonApplication changes language configurationNext responseRadiance

The underlying business workflow does not need to restart.

Language detection vs Code-Switch

These solvedifferent Problems

Automatic language detection answers which language is being spoken. Code-Switch recognition answerswhat happens when more than one language appears naturally in the same conversation.

Language detection

Primarily one language. Detect the language, then transcribe.

Code-Switch

Multiple languages mixed. Use the Code-Switch model for a continuous transcript.

Primarily one languageDetect languageTranscribe
Multiple languages mixedCode-Switch modelContinuous transcript

Applications should not treat these as identical features.

Real-world multilingual audio

Language is onlyone Challenge

A production voice agent also needs to handle:

  • Accents
  • Background noise
  • Phone compression
  • Multiple speakers
  • Fast speech
  • Interruptions
  • Names
  • Addresses
  • Numbers
  • Product terminology
  • Mixed-language vocabulary

A clean studio recording cannot represent all of these conditions. Production testing should include the audio users actually produce.

Telephony audio

Multilingual callers may still arrivethrough the phone Network

A multilingual voice experience must work beyond a browser microphone. Horizon supports telephony-oriented streaming audio.

Documented call audio

  • SIP / PSTN workflows
  • μ-law
  • A-law
  • 8 kHz telephony audio
Phone callTelephony audioHorizonAI applicationRadianceCaller

The language layer and telephony layer need to work together.

TTS audio formats

Generate speech for the channelwhere it will be Played

Radiance supports PCM, WAV, MP3, OGG Opus, FLAC, μ-law, and A-law.

Web

  • MP3
  • OGG
  • PCM

High-quality audio

  • WAV
  • FLAC

Telephony

  • μ-law
  • A-law

This allows multilingual generated speech to fit web, mobile, media, and telephone applications.

Speaking speed and expression

Language is not the onlypart of voice Quality

Radiance documents speech controls including speed from 0.25× to 4.0× and 11 expression styles for relevant workflows.

A voice agent may need different speaking behaviour for

  • Customer support
  • Notifications
  • Conversational assistants
  • Informational workflows

The application can configure the speech output according to the experience it is building.

Speaker diarization

Preserve multiple participants in multilingualRecordings

A recorded multilingual customer call may include two or more people. Speaker diarization can separate transcript segments according to speaker.

SPEAKER_00

I need to change my booking.

SPEAKER_01

Sure. Which date works for you?

Use for

  • Contact-center calls
  • Sales conversations
  • Support
  • Meetings
  • QA
  • Search
  • Training

This remains useful regardless of the language being spoken.

Word-level timestamps

Connect multilingual textback to the Audio

Verbose transcription can include:

  • Word start time
  • Word end time
  • Confidence information

Use for

  • Call review
  • Search
  • Training
  • Captions
  • Transcript playback
  • Analytics
SearchA phraseFind transcript occurrenceJump to audio

The same structured transcription workflow can support multilingual recordings.

Post-call intelligence

Turn completed multilingual conversationsinto structured Information

For supported batch workflows, Shunya supports optional fields such as:

  • Intent
  • Sentiment
  • Emotion
  • Summary
  • Speaker diarization
  • Speaker identification
  • Keyterms
  • Redaction
  • Word timestamps
Recorded conversationHorizonTranscriptIntelligenceCRM / search / analytics / LLM

These enrichments are batch capabilities: they run on the completed call, not live on every voice-agent turn.

Customer support

Serve customers withoutforcing one Language

A support voice agent can use the same application logic across different language experiences.

User A · English

Where is my order?

User B · Hindi

मेरा order कहाँ है?

User C · Hindi-English

Mera order dispatch ho gaya hai kya?

Speech recognitionSame order-support applicationAppropriate TTS response

The workflow remains customer support. Only the speech path changes.

Sales

Let prospects explain requirementsin their preferred Language

A prospect may explain the same requirement in English, Hindi, or a mixed-language conversation.

Prospect

We need transcription for our customer-support team.

Multilingual prospectHorizonSales / qualification applicationResponseRadiance

Sales logic does not have to be tied to a single spoken language.

Appointment booking

Book through multilingualConversation

A user may ask for the same appointment in more than one way.

User

I need an appointment Friday morning.

User

Friday morning appointment chahiye.

UserHorizonBooking applicationScheduling systemRadiance

The scheduling system remains the source of truth. The speech layer adapts to the conversation.

Banking and financial services

Financial conversations areoften Multilingual

Customers may discuss account actions using a mixture of local-language and English financial terminology.

Customers may discuss

  • Payment status
  • Account questions
  • Transactions
  • KYC
  • Service requests
  • Disputes
Customer speechHorizonFinancial applicationApproved workflowRadiance

The speech platform handles the language. The financial system handles the account action.

Healthcare

Multilingual speech also mattersin healthcare Workflows

Shunya separately documents Horizon Medical for healthcare-oriented speech recognition.

A healthcare voice application may need

  • General appointment conversations
  • Healthcare customer service
  • Clinical speech recognition
  • Medical terminology

General multilingual speech

Universal, Indic, or Code-Switch.

Specialised clinical speech

Appropriate Med workflow.

The use case should determine the speech model.

Expand to new languages without rebuilding the app

Separate the language layer from the businessLayer

Suppose the application first launches in English. Later it expands to Hindi, then Spanish, then Tamil.

Version 1

English, on the same business application.

Version 2

English + Hindi, on the same business application.

Version 3

More supported languages, on the same business application.

The speech layer can expand while the core business workflow remains intact. This is the architectural advantage of a multilingual speech platform.

Do not assume every language combination

A supported language and a supported Code-Switch pairare not always the Same

This distinction is important.

A language may be supported individually by Horizon Universal. That does not automatically prove that every possible mixed-language pair is available through the same Code-Switch route. Use Horizon Code-Switch where the required language combination is supported by the deployed configuration.

Check current deploymentConfirm language / pairChoose modelTest real audio

Do not hard-code assumptions about every possible language combination.

Check current language coverage

Use the live language list whenexact coverage Matters

Language coverage grows over time. For the exact languages you can deploy today,use Shunya's current language-listing API.

APIGET /languagesCurrent language namesISO codes and scriptsApplication configuration

This is more reliable than hard-coding a fixed language count into production logic.

Developers

Build one speech layer for multilingual voiceAI

InputUser audioHorizonChoose Universal, Indic, Code-Switch, or a domain model.TranscriptMachine-readable textAI / LLM / applicationConversation context. Business rules. Retrieval. Actions.Response textRadianceLanguage-specific / multilingual synthesisOutputSpoken response

Live

Use streaming STT and TTS for real-time voice agents.

Batch

Use completed recordings for post-call transcription and intelligence.

Language configuration

Use explicit language codes where known, automatic language detection where appropriate, and Code-Switch routes for supported mixed-language conversations.

Coverage

Use the current deployment's language listing when exact routable coverage matters.

Built for production multilingual voice AI

Beyond translatingone Sentence

A production multilingual voice agent may need:

  • Global STT
  • Indian-language STT
  • Code-switching
  • Automatic language detection
  • Streaming transcription
  • Streaming synthesis
  • Multilingual TTS
  • Cross-lingual synthesis
  • Mixed-language TTS
  • Telephony audio
  • Speaker diarization
  • Word timestamps
  • Post-call intelligence
  • API integration
  • Deployment-specific language routing
RequirementShunya speech stack
Global Multilingual STT
Indic STT
Code-Switch STT
Automatic Language Detection
Streaming STT
Batch STT
Multilingual TTS
Streaming TTS
Batch TTS
Cross-lingual Synthesis
Mixed-language TTS
Telephony Audio
Speaker Diarization
Word Timestamps
Conversation Intelligence
API Integration
UserHorizonAI / applicationRadianceUser

One conversational architecture. Multiple language experiences.

Questions

Frequently Asked Questions

What is a Multilingual AI Voice Agent?

A Multilingual AI Voice Agent is a conversational application that can understand and respond to users across multiple supported languages using speech recognition, AI/application logic, and speech synthesis.

Does Shunya support multilingual speech recognition?

Yes. Horizon supports broad multilingual recognition across global and Indian languages.

Which model handles international languages?

Horizon Universal is Shunya's broad global-language speech-recognition model.

Which model should I use for Indian languages?

Horizon Indic is designed for supported Indian languages and Indian speech patterns.

What is Horizon Code-Switch?

Horizon Code-Switch is designed for conversations where supported languages naturally mix within the same sentence or interaction.

Can Shunya handle Hinglish?

Yes. Hinglish is a core example of mixed-language conversational speech.

Can it handle Tamil-English and other mixed-language speech?

Code-Switch is designed for mixed-language conversations, but the exact supported language combination should be confirmed in the deployed configuration.

Does Shunya automatically detect language?

Yes. Horizon supports automatic language detection when language is not explicitly locked.

Should I use automatic detection for every call?

Not necessarily. When the expected language is known, explicitly locking it can reduce ambiguity, particularly for short audio.

Can one voice agent serve multiple languages?

Yes. The same business application can use different Shunya recognition and synthesis configurations according to the user's language.

Does Shunya support multilingual Text to Speech?

Yes. Radiance is available as a multilingual speech-synthesis platform.

Does Radiance support Indian languages?

Yes. Radiance has a named voice catalogue across English and major Indian languages.

Does Shunya support international-language TTS?

Radiance documents extended world-language synthesis, including languages beyond the core named Indic and English voice catalogue.

Does every supported language have named male and female voices?

No. Not every language has dedicated named voices. Use the voices available in the current deployment.

Can the same supported voice speak more than one language?

Radiance documents cross-lingual synthesis, subject to the capabilities available in the relevant account and deployment.

Can generated speech mix languages?

Yes. Radiance documents mixed-language synthesis, including Hindi and English conversational speech.

Can Shunya work in real time?

Yes. Horizon supports real-time streaming transcription and Radiance supports streaming speech synthesis.

How quickly can TTS begin speaking?

Radiance documents a first-audio streaming expectation of under 350 ms for relevant workflows.

Does Shunya support telephony?

Yes. Shunya's speech workflows support telephone-oriented audio including μ-law and A-law.

Can the agent switch language during a conversation?

The surrounding application can change language configuration based on the user's request or conversation state, subject to the supported speech workflows.

Can customer and agent speakers be separated?

Yes. Speaker diarization is available in supported STT workflows.

Can completed multilingual calls be summarised?

Yes. Summarisation is part of Shunya's supported batch conversation-intelligence layer.

Can intent and sentiment be analysed?

Yes. Intent and sentiment are supported as batch intelligence capabilities.

How many languages does Shunya support?

Horizon Universal recognises 204 languages and Horizon Indic 55 Indian languages; Radiance Universal speaks 669 languages and Radiance Indic 23 Indic languages. When exact deployable coverage matters, use the current Shunya GET /languages endpoint rather than a hard-coded number.

Does support for two languages mean they are automatically supported as a Code-Switch pair?

No. The relevant mixed-language combination should be verified against the deployed Code-Switch configuration.

Can I use one architecture for English, Hindi, Spanish, and other supported languages?

Yes. One speech platform can use different model and language routes according to the conversation.

Get started

Build voice AI for the way peopleactually Speak

Your users do not all speak one language. And multilingual users do not always stay inside one language. Use Horizon Universal for globalspeech. Use Horizon Indic for Indian-language conversations. Use Horizon Code-Switch for supported mixed-language interactions.Use Radiance to speak back through one multilingual voice platform. Test English, Hindi, regional languages, internationallanguages, Hinglish, code-switched speech, telephone audio, and the real conversations your users actually have.