Deepgram Flux TTS: Conversation-Native Text-to-Speech for Voice Agents, Architecture, Accuracy and Pricing

What Flux TTS is

Flux TTS is a text-to-speech model from Deepgram, launched on 12 August 2026 as the second model in its Flux family after Flux STT, its speech-to-text model for live conversations. Deepgram calls it the first conversation-native text-to-speech model built for real-time voice agents: instead of treating each line as a fresh request, it keeps the whole conversation in memory while it speaks. Deepgram argues that most text-to-speech was built for narration such as audiobooks and phone menus, so in a voice agent the tone resets every turn, developers must hand-tune delivery with SSML and style tags, and a line cannot be changed once it starts. On LMArena's leaderboards Flux TTS has no LMArena rating.

How it behaves in a conversation

Deepgram says Flux TTS adjusts tone, pacing and emotional register automatically, without SSML, style tags or prompt engineering, because it learned from real human conversation, including disfluencies and shifts in tone, rather than from studio recordings with style markup. Tone, pacing and pronunciation hold across a whole call. Each utterance has an explicit lifecycle of Idle, Generating and Completing, with events an agent can act on. When a caller interrupts, the server reports exactly what the caller heard so the agent can resume cleanly, and speed, pronunciation and pacing can be changed while the model is still speaking. Server-side flushing decides when a phrase is ready, so developers do not need their own chunking logic.

Architecture and latency

Deepgram names three design choices. A high-fidelity neural codec, from the same family as recent research by Kyutai, Meta and Google, avoids capping expressive range. Interleaved text-to-audio generation starts streaming audio on the first token, keeping time to first audio under 200 milliseconds regardless of response length, and as low as 80 milliseconds. A Mamba state-space backbone with fixed-size memory holds context across a whole session without the quadratic cost of transformers. Deepgram notes that callers start to notice lag at around 800 milliseconds of total round trip shared by speech recognition, the language model and speech synthesis.

Accuracy and deployment

Deepgram reports a median word error rate of 2.2% in its read-aloud tests, which it says is roughly half that of ElevenLabs and a third of Cartesia, and 3.4% on hard prompts such as account numbers, drug names, currency and dates. Flux TTS runs in the cloud, self-hosted or on-premises with the same model, and Deepgram says HIPAA, data residency and enterprise compliance are built in. Flux STT and Flux TTS run as a paired stack behind one API, and Deepgram plans for the two models to share state directly.

Pricing

Deepgram bills text-to-speech per 1,000 characters of input text. Flux TTS costs $0.045 per 1,000 characters on both the pay-as-you-go and Growth plans, against $0.030 and $0.027 for Aura-2 and $0.015 and $0.0135 for Aura-1, with Enterprise pricing by contract. The model was free until 12 September 2026, and Deepgram's pricing page offers $1 in credits for every $1 spent on Flux TTS, up to $500, through 31 December 2026.

Official Resources