Deepgram Flux TTS
Audio modelCard summarizedThe first conversation-native TTS built for real-time voice agents.
Source: Introducing Flux TTS (launch post) and Flux TTS developer docs (no model card published)
Where to call it
- Deepgram API (streaming)
flux-{voice}-en
Good for
- Turn-based voice agents, customer service bots and phone assistants
- Streaming LLM output to a speaker in real time with barge-in
- Entities voice agents must say: account numbers, drug names, dosages, prices, dates
Overview
Deepgram · 03TTS, front
Flux TTS
The first conversation-native TTS built for real-time voice agents.
Typical price
$0.045 / 1K chars
Flux TTS
- Max length
- 1 h per WebSocket session
- Languages
- English only
- Modes
- text-to-speech, streaming (WebSocket)
- Released
- 12 Aug 2026
Flux TTS is Deepgram's text-to-speech model for real-time voice agents, the second model in the Flux family after Flux STT. It keeps conversation state across turns so tone and pacing hold through a call, reports exactly what the caller heard when interrupted, and streams audio from the first token. Deepgram says it uses a Mamba state-space backbone and a high-fidelity neural codec, and ships at GA with English voices. launch post; Flux TTS overview
flux-haley-en- Typical price
- $0.045 / 1K chars
- Max length
- 1 h per WebSocket session
- Languages
- English only (American, British, Irish, Australian, Indian, Singaporean and Filipino accents)
- Weights
- Closed
- Released
- 12 Aug 2026
- text-to-speech
- streaming (WebSocket)
- batch (REST)
- turn-based synthesis with cross-turn context
- interruption handling
- mid-stream speed change
- expressivity (beta)