Thrifty AI × Cartesia: giving Sonic-3.6 voice agents a face, with Hinglish built in

Cartesia has become one of the default voice stacks for real-time agents, because it is fast and it sounds good. Many of the teams who talk to us already have a Cartesia voice agent on the phone and want to put it somewhere a phone can't go: a website, a meeting, a kiosk or a lobby. This post explains how that works with Thrifty, and the use cases where we have seen a face make the biggest difference.

What is Cartesia Sonic-3.6?

Sonic-3.6 is Cartesia's latest streaming text-to-speech model, released in August 2026. Cartesia reports under 90 ms latency and a generation speed of about 132 characters per second, roughly double its previous conversational model. It covers 44 languages and 61 locales, including 11 Indic locales, adds Odia and Urdu, and supports Hindi-English (Hinglish) code-switching. Cartesia also reports top positions on Artificial Analysis speech arenas at launch.

Cartesia built its original Sonic models on state space models, an architecture designed for fast, streaming generation. Alongside Sonic, Ink-2 is Cartesia's streaming speech recognition model for voice agents, with built-in turn detection and strong handling of numbers, dates and emails.

Why does Hinglish matter so much?

Most real conversations in India don't happen in one language. A customer might start in English, explain the problem in Hindi and read out an account number in English again. Voice agents that force a single language either mispronounce half the sentence or sound robotic switching between two voices.

Sonic-3.6's code-switching lets one voice handle that naturally. Pair it with a Thrifty AI human and you get a face that speaks the way your customers actually speak, which is why we see it chosen for Indian banking, telecom and citizen-services pilots.

People don't switch languages to be difficult. They switch because that's how they think. The AI should keep up.

How does a Cartesia voice agent get a face?

The flow is simple, and nothing about your Cartesia setup has to change:

• The user speaks; speech recognition (Ink-2 or another engine) transcribes and detects the end of the turn.
• Your LLM decides what to say and which tools to call.
• Sonic-3.6 streams the reply as audio.
• Thrifty's Maya face renders lip-synced video from that audio stream and delivers it over WebRTC, alongside Magic Canvas visuals if you use them.

When is a face worth adding to a voice agent?

Not every voice agent needs one. Voice-only is the right answer for a shipping-status check or a one-field confirmation. A face starts to pay for itself when the conversation is long, high-stakes or visual.

Cartesia or ElevenLabs?

We are often asked to pick. We think both are excellent, and the right answer depends on the job. Here is how we explain the trade-offs today; both vendors ship quickly, so revisit this when you start a new project.

How do you set it up in Thrifty Studios?

• Create or pick an AI human (stock face, one photo, or a consented clone).
• In the voice settings, choose Cartesia and select a Sonic-3.6 voice, or connect your existing Cartesia voice.
• Set the languages the AI human should handle, including Hindi-English if you serve Indian customers.
• Connect your LLM and knowledge base, then deploy to the web, Google Meet, a kiosk or your own servers.

Frequently asked questions