Google began rolling Personal Intelligence to Gemini on Android Auto this week. With permission, drivers can ask the assistant about travel plans or personal details drawn from Gmail and Google Photos while the phone stays in their pocket.
The same days brought two competing voice releases. Mistral shipped Voxtral 3, covering speech recognition and text-to-speech plus a voice-agent SDK, with support for more than 50 languages and deployment on a customer’s own infrastructure.
ElevenLabs introduced Eleven Realtime, a speech-to-speech family in Mini and Max versions. The company says it can begin speaking in about half a second and handles more than 90 languages.
The interfaces are converging
Text chat and voice used to be separate products with separate economics. They are now one stack: a recogniser, a reasoning model and a synthesiser, all measured by latency.
Latency is where differentiation sits. In a car, a two-second pause reads as a broken assistant. In a call centre, it reads as a bad line. Buyers of both workloads will pay for sub-second response.
Privacy follows the microphone
Personal Intelligence is opt-in and permissioned, but the trade is plain: a dashboard assistant that knows your calendar is also a device that can be asked about your calendar.
Voice models that run on-premises, like Voxtral’s self-hosted option, are the answer some regulated buyers will want.
