Perplexity in a Hyundai, Gemini in a Skoda: how generative-AI voice assistants actually work in cars sold in India

Three cars launched in India this October talk back in full sentences. The Hyundai Bayon ships with the company's first generative-AI voice assistant, built on Perplexity for answers and SoundHound AI for speech. The facelifted Skoda Slavia carries an "AI Companion" running Google Cloud's Automotive AI Agent, which is Gemini. And Mercedes-Benz has been rolling the same Google agent into its MBUX Virtual Assistant since the new CLA. This is the moment large language models moved from your phone into the dashboard of a ₹10 lakh SUV. Here is what is actually happening under the glass.

The three-layer stack

Every one of these systems is the same shape, whichever logo is on the bonnet.

Layer one: speech in, speech out. A wake word or a steering-wheel button opens the microphone. An automatic speech recognition engine turns your voice into text; a text-to-speech engine reads the answer back. In the Bayon this layer is SoundHound AI, a company whose core business is in-car voice. It is tuned for cabin noise, Indian English accents and interruptions. Skoda says its system is "designed to recognise Indian English accents", which is Google doing the same job with its own models.

Layer two: the brain. The text goes to a large language model that works out what you want and composes a reply. Hyundai uses Perplexity, which is an LLM wrapped around live web search, so a question like "what is the weather at Srisailam this weekend" is answered from current data rather than from the model's training. Skoda and Mercedes use Google Cloud's Automotive AI Agent, which Google describes as "built using Gemini with Vertex AI" and "specially tuned" for carmakers. Its party trick is multi-turn memory: ask for an Italian restaurant nearby, then "does it have good reviews", then "what is the most popular dish", and it keeps the thread. Google says it pulls from Google Maps Platform, which is updated with over 100 million edits a day.

Layer three: the car. Anything that touches the vehicle, such as climate, media, navigation routing, calls, window and sunroof control on some models, goes through a separate command layer that does not need the cloud. Hyundai is explicit about this split: the Bayon has more than 500 embedded voice commands in English, Hindi, Hinglish, Tamil, Telugu and Bengali that work with no data connection, while the generative assistant is English-only and online-only. Skoda lists music, calls and climate control as hands-free functions alongside "real-time information".

That split matters more in India than almost anywhere else. The cloud layer dies the moment you lose 4G on NH44 past Kurnool; the embedded layer keeps working.

Why now, and why in cheap cars

Two things changed. First, the cost of inference collapsed. Perplexity and Google sell these capabilities as a subscription to the carmaker, so the per-car cost is small enough to put in a mass-market SUV rather than an S-Class. Second, the hardware in the dashboard got serious. Hyundai's ccNC cockpit in the Bayon is accelerated by NVIDIA; Google's agent is designed to run hybrid, with Qualcomm's Snapdragon Digital Chassis handling on-device work and the cloud handling the heavy reasoning. A head unit that can render two 12.3-inch displays with navigation and camera feeds has the headroom to run a speech pipeline locally.

The third enabler is over-the-air updating. Hyundai says the Bayon has up to 19 controllers that accept OTA updates, "enabling remote feature rollouts" without a dealer visit. That is the real story: the assistant you get on delivery day is version one. Google's own Gemini-in-cars post shows how fast the roadmap moves, with Gemini rolling out across the 250 million cars that support Android Auto and the 50-plus models with Google built-in. If your car can take OTA updates, your assistant will get smarter; if it cannot, it will be frozen at launch spec.

What it is good at, and what it is not

Good: trip planning ("plan a two-day route to Hampi with a lunch stop"), live information (weather, scores, news, fuel prices where the data exists), places with follow-up questions, and dictating messages in natural language rather than a fixed template. Google's example of telling Gemini to "always send messages to Joe in Spanish" and having it draft rather than transcribe is the kind of thing that genuinely saves attention at the wheel.

Not good, and not designed for: anything safety-critical. None of these assistants touches steering, braking, ADAS thresholds or drive modes. Those live on separate, safety-rated electronic control units and the language model has no path to them. Hyundai's press footnote says it plainly: "AI generated responses are for informational and convenience purposes only and are not a substitute for human judgment." Treat the assistant as a well-read passenger, not a co-driver.

Also not good yet: Indian languages in the generative layer. The Bayon's Perplexity assistant is English-only at launch, even though its embedded commands cover five Indian languages. Skoda's Google agent supports multiple languages globally but Skoda's India release only commits to Indian English accents. Expect Hindi and Telugu generative support to arrive by OTA rather than at launch.

What to check at the dealership

Ask which trim gets the assistant. In the Slavia it starts at the Signature trim, not the Classic. In the Bayon it comes with the ccNC head unit on higher variants. Ask what the connected-services subscription costs after the free period; Hyundai bundles three years of Bluelink with the Bayon. Ask whether the assistant works with the car's own SIM or needs your phone's hotspot. And ask who holds your voice data: the carmaker's terms, Perplexity's or Google's. The honest answer at most dealerships today is "we are not sure", which tells you the industry is still catching up with its own feature list.

The Auto Grid angle

A software-defined car updates itself over the air, remembers your restaurant preferences and reroutes around a jam before you have seen it. It still cannot do anything about a stone chip on the bonnet. The irony of the Bayon, Slavia and CLA is that the parts of the car that get cleverer every month sit behind the parts that get worse every month: painted steel and plastic that take the ORR's gravel, two-wheeler mirrors and underground-parking pillars. The radar and camera hardware that these software platforms depend on also lives in the front bumper and behind the windscreen, so a bumper respray after a minor hit can mean a sensor recalibration. PPF on the front end keeps the paint original, keeps the sensor mounts untouched, and means the only thing that changes on the car over five years is the software.

Where this goes

Google's Automotive AI Agent and Perplexity's automotive push are both platform plays: the carmaker buys the brain, keeps the branding. That means the next wave, including Maruti, Tata and Mahindra, will arrive fast and with the same two or three vendors underneath. The differentiators will be the boring ones: how many Indian languages the generative layer speaks, how much works offline, and how long the carmaker keeps paying for updates after you have paid for the car.