Google is bringing conversational Gemini support in Google Maps to Apple CarPlay while also expanding personalization in Android Auto. This matters if you want to use your voice for navigation, appointments, or other information while traveling. At the same time, new real-time speech models from ElevenLabs and Mistral AI show why digital assistants could start feeling less like voice menus and more like conversations.
What is changing in the car
Gemini is a form of Artificial Intelligence (AI) that can understand spoken requests, combine information, and formulate replies. In Google Maps on Apple CarPlay, you can start the assistant by tapping an icon on the infotainment display. According to the report on the CarPlay launch, users in the United States can give it requests with multiple steps.
You could first ask for recommendations on places to visit, then tell Gemini to guide you to one while avoiding toll roads. The assistant is also intended to help find stops and report traffic disruptions. These conversations can replace several separate inputs, but they do not eliminate the risk that AI will misunderstand you or provide inaccurate information.
Google also says a global CarPlay rollout is underway, though the report gives no specific countries or dates. That means availability in Switzerland cannot yet be confirmed from the supplied information. The CarPlay announcement also does not mention the Personal Intelligence features announced for Android Auto.
Personal Intelligence is Google’s optional system for letting Gemini use personal content from its services. In Android Auto and vehicles with Google built-in, meaning Google software integrated directly into the infotainment system, Gemini can draw from Gmail, Calendar, Maps, Photos, YouTube, and Search. As the coverage of in-car personalization explains, this connection is only supposed to be used if you enable Personal Intelligence.
What real-time models improve
The second change is happening not in the vehicle itself but in the underlying speech models. A real-time speech model processes spoken language with little enough delay to support a conversation without long pauses. Speech-to-speech models accept audio and produce a spoken response rather than merely displaying text.
ElevenLabs has introduced two such models under the Eleven Realtime name. According to the company, the more capable v1 Max responds in 0.57 seconds and scores 98.8 percent on Big Bench Audio, a standardized test of audio capabilities. Those figures were not independently verified in the supplied report, and the claim that no competing model is both faster and more accurate also comes from the provider.
The smaller v1 Mini is intended for natural conversations at lower cost, according to the report on Eleven Realtime. It is designed to recognize whether you have finished speaking or whether a brief “mm-hmm” simply means you are still listening. V1 Max is aimed at more complex tasks such as customer support.
Mistral AI is taking a more modular approach with Voxtral 3. The system combines speech-to-text, which converts spoken language into written text, with text-to-speech for audio output and a software kit for building voice agents. According to the provider, these parts can bring listening, processing, access to external tools, and speaking together in one continuous stream.
Transcription supports more than 50 languages, according to the Voxtral 3 announcement report. The speech output is said to cover more than 35 languages and 50 dialects, with emotional adaptation. The source does not specify the quality available for Switzerland’s individual national languages or regional varieties.
A separate short report describes Voxtral 3 as a family of three transcription models, one text-to-speech model, and a LiveKit-based voice agent software kit. It says Voxtral Small Transcribe 3 has an error rate of 2.1 percent on the Artificial Analysis test and ranks third. Because the item is brief and offers no independent assessment, these Voxtral 3 benchmark figures should not be treated as proof of quality in every conversation.
Who benefits from it
In the car, the most immediate benefit is fewer interactions with the screen. If your flight details are stored in a confirmation email, Gemini is supposed to retrieve the flight status while you drive to the airport. When you say, “Take me to my doctor’s appointment,” the assistant can, according to Google, find the address in a connected calendar and start navigation.
Another example from Google combines image search with navigation: Gemini is supposed to locate the car wash shown in a photo you took the previous week and guide you there. That is more convenient than searching emails, calendar entries, or photos while driving. The convenience, however, comes precisely from connecting several sources of personal data.
Outside the car, real-time models could make customer service conversations feel more natural. Mistral describes voice agents that can book appointments or check invoices during a call. That may interest businesses because a single conversation can combine voice input, data retrieval, and spoken output; for you, the key questions are whether it is clear that you are speaking to AI and what happens to your data.
The vehicle platform matters as well. Android Auto projects an interface from a connected phone, while Google built-in runs directly in the vehicle’s infotainment system. Ars Technica notes that relatively few vehicles have Google built-in and that new features do not necessarily arrive there first.
Pros and Cons of voice AI on the road
Pros:
- Fewer manual steps – Multi-step requests can combine searching, selecting, and navigation in one spoken exchange.
- Personal context – Calendars, emails, and photos can produce more relevant responses without requiring you to repeat every detail.
- Smoother conversations – Lower latency and better detection of pauses can make an assistant respond more naturally.
- Broader language coverage – Voxtral 3 claims support for many languages and dialects, although quality by language remains unclear.
Cons:
- More personal data – The benefits of personalization depend on access to sensitive material such as appointments, messages, and photos.
- Uncertain availability – Your platform, region, vehicle, and account determine which features you actually receive.
- Errors at the wrong time – Misunderstandings or fabricated answers can be more distracting on the road than during quiet desk use.
- Hard-to-compare metrics – Some latency and benchmark numbers come from providers and reveal little about accents, background noise, or real conversations.
The cost comparison is incomplete as well. ElevenLabs bills Eleven Realtime by the minute plus token charges; tokens are small processing units used for language or audio. V1 Max costs more than v1 Mini, but the supplied source does not include exact amounts.
Mistral says transcription starts at $0.0035 per minute. That may represent only part of the total cost if speech output, external services, or a complete voice agent also needs to be operated. The supplied reports do not state a price for Gemini’s in-car features.
What this means for you
If you are starting out, first check your platform, region, and enabled settings. On CarPlay, you need the relevant Gemini feature in Google Maps; personalized answers in Android Auto require Personal Intelligence to be available and switched on. Begin with a simple, low-risk task such as finding a stop rather than immediately connecting email, calendar, and photo content.
More advanced users can compare how reliably multi-step requests work and how the assistant handles road noise, interruptions, or ambiguous place names. For professional use, it also makes sense to evaluate transcription, speech output, response time, and data hosting separately. Low latency alone does not create a useful assistant.
For Switzerland, the current position is unusually clear but not especially satisfying. According to the report on regional restrictions, Personal Intelligence for Android Auto is not initially rolling out in the European Union, Switzerland, the United Kingdom, or Nigeria. There is no stated date for wider access, and the broadly worded global CarPlay rollout does not resolve that uncertainty.
On privacy, Mistral says Voxtral 3 can run on an organization’s own infrastructure or through application programming interfaces hosted in the European Union. An application programming interface connects the speech service with other software. This can offer organizations more control, but it does not replace checking which conversation data is stored, shared, or used for other purposes.
Voice AI in cars is becoming more practical because it can understand longer requests, use personal context, and respond faster. Its real value still depends more on region, platform, language quality, and data permissions than polished demonstrations suggest. The unresolved risk is that the most convenient assistants need the most context while remaining capable of mistakes in situations where your attention belongs on the road.
Sources
- Conversational Gemini navigation comes to Apple CarPlay – The Verge, 2026-10-21
- Google’s more personalized Gemini AI is heading to your car – The Verge, 2026-10-20
- Gemini’s Personal Intelligence is coming to cars – Ars Technica, 2026-10-20
- Google rollt Personal Intelligence für Android Auto aus: Warum deutsche Autofahrer leer ausgehen – t3n, 2026-10-21
- ElevenLabs stellt Eleven Realtime vor: Speech-to-Speech-Modelle mit unter einer Sekunde Latenz – The Decoder, 2026-10-20
- Mistral AI stellt Voxtral 3 vor: Neue Sprachmodelle sollen Echtzeit-Sprachagenten ermöglichen – The Decoder, 2026-10-20
- Mistral AI Releases Voxtral 3: Speech-to-Text, TTS and a Voice Agent SDK – MarkTechPost, 2026-10-21
Image: Dolores Reyes via Pexels