A voice agent that gets the details right

Flux Text-to-Speech (TTS) is a real-time conversation-focused voice agent tool. Unlike other TTS models, Flux TTS reads the entire conversation, maintaining context from the very first response to the last one. Currently, developers can test Flux TTS free of charge, with up to 45 concurrent streaming connections worldwide (5 connections available in EU/AU). Standard pricing begins on September 13, 2026.

Flux TTS was ranked through a blind two-alternative forced choice methodology, comparing 14,400 paired comparisons across 12 models and 73 voices over 100 scripts derived from real-world use cases. The ranking was based on each model's best voice and the ability to sound more natural or expressive. The technology integrates seamlessly into existing voice-agent stacks, whether running in the cloud, VPC, or on-premises.

Audio generated by Flux TTS includes an inaudible watermark, developed by Deepmark, to indicate that the audio was created using Deepgram technology. This watermark remains intact during live, real-time conversations, ensuring the authenticity of the voice AI. Deepgram believes that voice will become the cornerstone of AI experiences and aims to provide trust in conversation through Flux TTS.

Flux TTS aims to address the challenges developers face when building voice agents. Traditional TTS models often produce speech that deteriorates over a conversation, requiring extensive SSML tuning to maintain a natural tone and expressiveness. Flux TTS combines conversation-aware speech with low-latency performance and flexible deployment options, offering a solution that improves the caller experience without the need for line-by-line tuning or relinquishing control over the existing stack.

Coval, a company that builds voice agents, has found that the TTS side of voice AI presents consistent failure modes, such as speech flattening after a few minutes, requiring heavy SSML tuning, and inconsistency across turns, which disrupts the flow of a real conversation. Deepgram's Flux TTS model directly tackles these issues by treating the conversation as the unit rather than individual lines.

This approach provides TTS models that can maintain high quality throughout a conversation, giving voice agent teams greater confidence in performance during rigorous, scenario-based evaluations.

By integrating Flux TTS into LiveKit's agent platform, developers can add conversation-aware voice to their existing real-time infrastructure. Alexis, a voice model developed with Flux TTS, has been described as human, expressive, and distinctly friendly—qualities that are particularly important in telehealth preconsult calls. Companies have reported significantly improved voice quality using Flux TTS, making it a strong contender for future voice-agent development. To switch to Flux TTS, developers simply need to change the model parameter to flux-{voice}-en.