Deepgram has expanded its Flux family with the launch of Flux TTS, a text-to-speech model specifically engineered for real-time voice agents. According to audioXpress, the San Francisco-based company describes this new model as conversation-native because it treats an ongoing exchange as the operating context for generation rather than focusing on isolated sentences. This approach extends the conversational architecture previously introduced in Deepgram speech-to-text technology to the realm of generated speech. The model retains conversational state between turns, allowing tone, pacing, and emotional register to remain consistent throughout longer interactions without requiring developers to use complex style tags or additional prompt engineering.
AudioXpress reports that the system manages explicit lifecycles for each conversational turn, enabling the software to return data identifying which portions of a response were rendered if an interruption occurs. This capability helps keep internal agent history synchronized with actual listener experience during critical tasks like account management or scheduling. Deepgram says the model can begin responding in as little as 80ms, including under production load, and accepts streaming output from large language models so generation starts before a full response is complete. While currently offering conversational English voices, the company plans to add more languages and voice cloning features later. The tool is generally available now with free access offered through September 12, 2026, before standard pay-as-you-go pricing begins on September 13, 2026, notes audioXpress.
For more information, visit audioxpress.com.