Deepgram Flux vs Aura-2: Which Text-to-Speech Should You Use?

Choose Deepgram Flux for real-time voice agents that hold a back-and-forth conversation. Choose Aura-2 for simpler jobs like reading IVR prompts, notifications, or short narration. Both are text-to-speech (TTS) models from Deepgram, so this is not a fight between rivals. It is a fit question.

First, clear up a common mix-up. Many people search "Deepgram Flux vs Nova" and think Nova is a TTS option. It is not. Nova-3 is Deepgram's speech-to-text (STT) model. It listens and transcribes the caller. Flux and Aura-2 do the opposite job: they speak. So Flux vs Nova is a category confusion, not a real choice.

This page explains the difference in plain terms. We cover what each model is built for, when the upgrade to Flux is worth it, and when Aura-2 is plenty. We also compare cost, latency, accuracy, and how you deploy them. By the end, you should know which model fits your voice project and why the Nova mix-up trips so many people up.

Deepgram Flux vs. Deepgram Aura-2: Side-by-Side

First, Is "Flux vs Nova" Even a Real Question?

No. Flux vs Nova is a category mix-up, and it is worth fixing before you compare anything.

Nova-3 is Deepgram's speech-to-text model. Its job is to listen to the person on the call and turn their words into text. Flux and Aura-2 are text-to-speech models. Their job is to speak back. One transcribes, the others talk.

In a voice agent, these layers work together, not against each other. The caller speaks, Nova-3 transcribes it, a language model decides what to say, and a TTS model like Flux or Aura-2 says it out loud. So the real question is never Flux vs Nova. It is Flux vs Aura-2 for the speaking role, with Nova-3 handling the listening role beside it.

What Is Deepgram Flux Built For?

Flux is Deepgram's conversation-native TTS, purpose-built for real-time voice agents.

Unlike a line-by-line voice, Flux reads the whole conversation as it unfolds. That means tone, pacing, and emotional register stay steady across the full session. You avoid the common failure where a voice sounds warm on pickup and flat by the third turn.

Flux also handles the hard parts of live talk. It has native interruption handling, so a caller can cut in and the agent responds naturally. Its time-to-first-audio can be as low as 80ms, which keeps replies feeling instant. It reads alphanumerics, drug names, and technical strings accurately, and it needs no SSML or prompt tricks to sound expressive.

• Reads the whole session, not just the current line
• Time-to-first-audio as low as 80ms
• Native interruption and turn-taking
• Accurate on alphanumerics, drug names, and technical strings
• No SSML or style tags required

What Is Deepgram Aura-2 Built For?

Aura-2 is Deepgram's fast, general-purpose TTS model. It is a strong fit for straightforward speaking jobs.

Think IVR prompts, order confirmations, appointment reminders, notifications, and simple narration. In these jobs the script is mostly fixed and the voice does not need to track a live back-and-forth. Aura-2 speaks the text cleanly and quickly.

Aura-2 is listed at $0.030 per 1,000 characters, which is roughly two to three cents per spoken minute. It is Flux's practical predecessor for the voice-agent use case, and for many simpler jobs it is all you need. Confirm the current rate at deepgram.com/pricing.

When Is Upgrading to Flux Worth It?

Upgrade to Flux when the voice has to hold a real conversation, not just read a script.

If callers interrupt, ask follow-up questions, or expect a natural back-and-forth, conversation-awareness matters. Flux keeps tone consistent turn after turn and handles interruptions without sounding robotic. That is exactly where a line-by-line voice starts to feel off.

When we evaluate TTS for a client's voice agent at Layer3 Labs, the failure we see most is a voice that sounds warm on pickup and flat by the third turn. That is the exact problem a conversation-native model targets. If your agent is a receptionist, a support line, or a booking assistant, Flux is likely worth it.

When Is Aura-2 Enough?

Aura-2 is enough when the speaking job is one-directional and the script is predictable.

You do not need conversation-awareness to read a confirmation number, announce a queue position, or narrate a fixed prompt. In those cases Aura-2 gives you fast, clear speech at a published rate you can budget against.

Aura-2 is also a sensible default when you are testing an idea or keeping costs tight on simple flows. You can start there and move to Flux later if the agent grows into live conversation. Both models deploy anywhere, so the migration path stays open.

How Do Flux and Aura-2 Compare on Accuracy?

Both models produce clean, natural speech, but Flux is tuned for the strings that break voice agents in production.

Flux reads alphanumerics, drug names, and technical strings accurately. That matters when an agent reads back an order number, a booking code, or a medication. A small stumble there can cost the whole call, so accuracy on these strings is a real differentiator.

Aura-2 handles everyday scripts well, which is why it fits reminders, prompts, and narration. For those jobs the words are common and predictable, so the accuracy bar is easier to clear. If your scripts are heavy with codes or clinical terms, lean toward Flux and test with your own examples.

How Do Cost, Latency, and Deployment Compare?

On cost, Aura-2 gives you a clear anchor at $0.030 per 1,000 characters, which is roughly two to three cents per spoken minute. That makes it easy to model a monthly bill from your expected volume. Flux launched free through September 12 as a promo, so verify its live pricing at deepgram.com/pricing before you plan a budget, since a launch rate can change.

On latency, Flux states a time-to-first-audio as low as 80ms, and that number is the one that keeps live replies feeling instant. Every layer in a voice agent adds delay, so shaving the time before the first word matters. Aura-2 is fast too, but Deepgram does not publish the same figure, so confirm it on the pricing and docs pages if latency is critical to your build.

On deployment, the two are even. Both Flux and Aura-2 can run in your own cloud or on-prem. That helps you meet data-residency, security, and compliance needs, which matters a lot in healthcare, finance, and other regulated work. So your choice between them comes down to the job and the budget, not where the model can run.

→ Compare per-token pricing for every AI model

The Verdict

The bottom line: pick the model by the job, not by the name. Flux and Aura-2 are both Deepgram TTS models, and Nova-3 is a separate STT model that listens rather than speaks. So route by workload.

Choose Deepgram Flux for real-time voice agents that hold a conversation, such as receptionists, support lines, and booking assistants. In those jobs, interruptions, steady tone across turns, and 80ms responses decide whether the call feels human. Choose Aura-2 for one-directional speaking jobs like IVR prompts, reminders, and simple narration, where its $0.030 per 1,000 characters rate is easy to budget and conversation-awareness is not needed. Many teams end up running both, using Aura-2 for fixed messages and Flux for live calls.

If you are not sure which side your use case sits on, map the call flow first. Count how often the caller talks back. Heavy back-and-forth points to Flux; mostly fixed scripts point to Aura-2. Always confirm current pricing at deepgram.com/pricing, since Flux is launch-fresh.