Early access · Singlish first

Voice AI that speaks like Southeast Asia.

Speech recognition, natural Singapore voices and a voice layer for agents, built for Singlish and code-switching and hosted in Singapore.

Request early access See the products

A real call, the way Singapore talks

Eh hi, can check if Dr Tan free 明天 afternoon or not?

Can, let me check for you ya…

Dr Tan is free at three o'clock tomorrow. Shall I book that?

Illustrative conversation.

Products

Two ways to build with Tavia.

VOICE AGENTS SDK · EARLY ACCESS

Bring your agent. We give it a voice.

Connect your prompt, tools and LLM. Tavia handles listening, turn-taking, interruptions and a fast talker that replies while your tools run, over phone or web.

Streaming speech recognition for Singlish and code-switching
Natural turn-taking and backchannels ("mm", "ya")
Fast tool routing, with read and write tools kept apart
Singapore voice out, any phone or web channel

const agent = tavia.agent({
  instructions: "Friendly clinic receptionist",
  voice: "sg-calm-1",
  tools: [findNearestBranch, bookAppointment],
});
await agent.connect(phoneCall);

API preview, subject to change.

TEXT TO SPEECH API · IN DEVELOPMENT

Voices that sound like home.

Calm, natural Singapore-accent voices for phone lines, apps and content, with real-time streaming or batch.

Switches between English, Mandarin and Malay mid-sentence
Custom pronunciation for local names, streets and brands
Pace, style and emotion controls
Instant and professional voice cloning, with consent

Languages

Singlish first. The region next.

Most voice AI is trained on American English. We start where it struggles most: the way Singaporeans really talk, switching languages in one breath.

FIRST · EARLY ACCESSSinglish & SG EnglishIncluding particles like lah, ya, can

FIRST · EARLY ACCESSMandarin + code-switchingEnglish and Mandarin in one sentence

IN DEVELOPMENTMalay · IndonesianFor Malaysia and Indonesia

ROADMAPThai · Vietnamese · TagalogAcross Southeast Asia

Why Tavia

Built here, for here.

Data stays in Singapore

Built to run in Singapore, which matters for clinics, finance and anyone under PDPA.

Feels like a real conversation

Designed for fast replies, natural "mm"s and clean interruptions, so calls flow like talking to a person.

Our own models

We train our own speech recognition, voices and classifiers for the region, not a wrapper around someone else's.

Safe by design

Voices are checked so they never copy a real person, and cloning requires consent.

Building a voice product for Southeast Asia?

We're working with a small group of early partners. Tell us what you're building.

Request early access