Deepgram launched Flux TTS, the second model in its Flux family, describing it as the first conversation-native text-to-speech model built for real-time voice agents; it follows Flux STT, which reworked speech recognition for live conversations, and is free until 12 September.
The diagnosis is more specific than the product claim. Today's text-to-speech was built for narration — audiobooks, voiceover, IVR prompts — and Deepgram's argument is that dropping a narration model into a live conversation breaks in identifiable ways. The first is tone: because each line is a fresh request and the model does not know it is inside a conversation, an empathetic tone cannot hold across a long call and a shift in the customer's mood cannot carry over. The failure is architectural rather than a quality problem, which is why tuning a voice does not fix a call that drifts by minute three.
Treating the turn as state rather than a request is the same separation described in GPT-Live splitting the conversation loop from deep reasoning, and it belongs with the latency-shaped design decisions collected in the agents hub.
This is a launch post: it names the failure modes and the promotional window without publishing comparative latency or listener evaluations. The condition to watch is whether Deepgram publishes turn-level measurements against a narration-trained baseline, because conversation-native is a claim about state handling that only a multi-turn test can show.