Deepgram’s 80ms Flux TTS targets real customer calls

Voice AI demos usually sound magical until a real customer interrupts, changes an order halfway through, or rattles off a phone number. Deepgram is tackling that gap with Flux TTS, a new text-to-speech model built specifically for live conversations.

Built for interruptions and real-world calls

Flux delivers response times as low as 80ms while maintaining conversational nuance. Rather than treating each turn as an isolated line of text, the model tracks context across the entire dialogue and adapts mid-call.

According to Deepgram, Flux is tuned to handle the messy dynamics of real speech that trip up traditional TTS, such as handling customer interruptions, pivoting when an order changes, and reading back fast number sequences accurately. The goal is to move voice agents out of sandbox demos and into actual enterprise phone calls. Developers can build with Flux TTS for free through September 12th.

Why it matters

Voice agents are moving into production customer service, where raw speed and conversational resilience matter far more than polished studio fidelity. An 80ms latency effectively eliminates the awkward dead air that tips users off that they are talking to a bot.

By focusing on unscripted user behavior like mid-sentence interruptions and rapid-fire numbers, Deepgram is attacking the specific failure points that cause people to hang up on automated agents. We think this sets a higher baseline for what real-time voice infrastructure needs to deliver, and the free window through September 12th makes it an easy benchmark test for developers building interactive call workflows.