xAI has launched its flagship voice model, Grok Voice Think Fast 1.0, designed for natural, multi-turn voice conversations. The company touts it as a tool for customer support and telemarketing.
https://twitter.com/xai/status/2047441173569216721
The model boasts top-tier intelligence and low response latency, promising fluid interaction even through complex, back-and-forth voice exchanges.
You can try it yourself by typing text at the demo site — it will respond with spoken audio. I gave it a spin: it handles Japanese with remarkably accurate accent, though I noticed one spot where the transcribed text didn't match the audio (it wrote "1885" when it probably should have been something else).
Grok Voice Think Fast 1.0 topped the τ-voice Bench, a benchmark that evaluates voice agents under realistic conditions like noise, heavy accents, and interruptions. The τ-voice Bench results show Grok Voice Think Fast 1.0 far ahead of competitors, as detailed in our separate article on xAI's new voice agent crushes rivals on real-world benchmark.
xAI says the model was "rigorously tested under the most challenging real-world conditions — phone audio, background noise, strong accents, frequent interruptions." It natively supports over 25 languages, making it suitable for global deployment.
The model is also good at extracting details like addresses, phone numbers, names, and account numbers even when the speaker talks fast or corrects themselves. It can perform background reasoning, so it handles complex queries and workflows in real time without slowing responses.
Grok Voice Think Fast 1.0 is already powering Starlink's customer support and sales. According to xAI, one in five inquiries that resulted in a purchase came from an AI conversation, and 70% of customer support issues are resolved by the AI.