# Migrating from Nova-3 to Flux

## Key Benefits of Flux

- **Model-integrated turn detection** (`StartOfTurn`, `EagerEndOfTurn`, `TurnResumed`, `EndOfTurn`)
- **Ultra-low latency** \~260ms end-of-turn detection (p50 at defaults)
- **EagerEndOfTurn events** let you **start LLM responses early**
- **Turn-based transcripts** for clean agent logic
- **Same Nova 3 transcription quality**
- **Simplified development** one API replaces complex STT+VAD+endpointing pipelines, and conversation-native events.
- **High configurability** - Configurable end-of-turn detection sensitivity, eager response thresholds, and turn-taking dynamics for optimized conversational flow

## Audio Requirements

- **Encoding:** See [Audio Format Requirements](#audio-format-requirements) table below
- **Sample rates:** See [Audio Format Requirements](#audio-format-requirements) table below
- **Channels:** Mono only
- **Chunk size:** **80ms strongly recommended** for optimal model performance and latency.

### Audio Format Requirements

| Audio Type    | Encoding                                                    | Container | `encoding` param | `sample_rate` param | Supported Sample Rates                     |
| ------------- | ----------------------------------------------------------- | --------- | ---------------- | ------------------- | ------------------------------------------ |
| Raw           | `linear16`, `linear32`, `mulaw`, `alaw`, `opus`, `ogg-opus` | None      | **Required**     | **Required**        | `8000`, `16000`, `24000`, `44100`, `48000` |
| Containerized | `linear16`                                                  | WAV       | **Omit**         | **Omit**            | Auto-detected from container               |
| Containerized | `opus`                                                      | Ogg       | **Omit**         | **Omit**            | Auto-detected from container               |
| Containerized | `opus`                                                      | WebM      | **Omit**         | **Omit**            | Auto-detected from container               |

## Migrating from Nova 3 to Flux

This guide will help you migrate from Nova 3 to Flux by highlighting key differences, setup changes, and implementation patterns.

### Differences

| Nova 3                                             | Flux                                        |
| -------------------------------------------------- | ------------------------------------------- |
| Streams transcripts continuously                   | Emits structured turn events                |
| Requires custom logic for barge-in and turn-taking | Has built-in turn state machine             |
| Returns transcripts only                           | Returns conversation events and transcripts |
| Designed for general real-time transcription       | Designed for conversational voice agents    |
| Focuses on accuracy and speed                      | Focuses on accuracy and turn awareness      |

#### Endpoint Usage

**Nova 3:**

Uses the listen v1 endpoint with the `nova-3` model option.

```
wss://api.deepgram.com/v1/listen?model=nova-3
```

**Flux:**

Uses the listen v2 endpoint with the `flux-general-en` model option.

```
wss://api.deepgram.com/v2/listen?model=flux-general-en
```

### Response Message Structure

#### Nova 3

```json
{
  "type": "Results",
  "channel": "transcript",
  "alternatives": [...]
}
```

#### Flux

```json
{
  "type": "TurnInfo",
  "request_id": "2ba892a1-6c0d-4d92-9b89-0000000000",
  "event": "Update",
  "turn_index": 0,
  "audio_window_start": 0,
  "audio_window_end": 0.47999996,
  "transcript": "",
  "words": [...],
  "end_of_turn_confidence": 0.0009,
  "sequence_id": 2
}
```

In addition to the transcript, flux responses include the:

- `event` field for turn-state changes
- `turn_index` to track turn lifecycle
- `audio_window_start` and `audio_window_end` to track the audio window.
- `end_of_turn_confidence` to track the confidence of the end of turn.
- `sequence_id` to track the sequence id of the messages.
- `words` array with word-level `start` and `end` timestamps (type `double`) on each word object, along with `word` and `confidence`.

### Implementation Pattern Changes

#### Nova 3 Approach

> Requires custom logic for barge-in and turn-taking.

- Send audio
- Receive streaming partial transcripts
- Decide when to interrupt your agent manually

#### Flux Approach

> Listens for structured events and removes the need for custom VAD or barge-in logic.

- `StartOfTurn`: Interrupt agent if it’s speaking
- `EagerEndOfTurn`: Medium-confidence end → start LLM reply
- `TurnResumed`: User kept talking → cancel reply
- `EndOfTurn`: High-confidence end → send transcript to LLM

By default, Flux only emits `Update`, `StartOfTurn`, and `EndOfTurn`.

### Simple Approach: Enabling End of Turn

:::callout{intent="info"}
For more information on using Flux with EndOfTurn only see the [Flux Getting Started Guide](/guides/streaming-audio-flux-quickstart)
:::

This is a simple approach using only `EndOfTurn` (lower latency, less complex, less LLM calls).

To enable end of turn use the `eot_threshold` parameter which allows for a confidence of (0.5–1.0) for `EndOfTurn` events.

#### Example

```text
wss://api.deepgram.com/v2/listen?model=flux-general-en&sample_rate=16000&encoding=linear16&eot_threshold=0.8
```

### Optimized Approach: Enabling EagerEndOfTurn + EndOfTurn

This is an optimized approach using both EagerEndOfTurn and EndOfTurn (lower latency, slightly more complex, more LLM calls)

To enable eager end of turn use the `eager_eot_threshold` parameter which allows for a Confidence of (0.3–0.9). You can also set the `eot_threshold` with a confidence of (0.5–1.0) to handle `EndOfTurn` events and use the `eot_timeout_ms` which defaults to 5000 ms to force a timeout after a specified time.

#### Example

```text
wss://api.deepgram.com/v2/listen?model=flux-general-en&sample_rate=16000&encoding=linear16&eager_eot_threshold=0.6&eot_threshold=0.8&eot_timeout_ms=7000
```

### Keeping Your Own Turn Detection

If you already have a turn detection stack you want to keep (VAD, endpointing, or push-to-talk), you don't have to adopt Flux's native detection. Set `eot_threshold=1.0` to suppress natural end-of-turn, then send a [`ForceEndTurn`](/guides/streaming-audio-flux-force-end-turn) message when your own detector fires. See [Bring Your Own Turn Detection](/guides/streaming-audio-flux-own-turn-detection) for the full recipe.

#### Example

```text
wss://api.deepgram.com/v2/listen?model=flux-general-en&sample_rate=16000&encoding=linear16&eot_threshold=1.0
```

### Nova 3 Migration Checklist

- [ ] Update WebSocket endpoint to `/v2/listen`
- [ ] Set `model=flux-general-en` and `encoding=linear16`
- [ ] Adjust client to parse `TurnInfo` messages
- [ ] Implement turn event handling (start, eager end of turn, turn resumed, end)
- [ ] Tune `eager_eot_threshold` and `eot_threshold` for your use case
- [ ] Remove custom VAD/barge-in logic (Flux handles this natively) — or keep it and drive turns with [`ForceEndTurn`](/guides/streaming-audio-flux-force-end-turn); see [Bring Your Own Turn Detection](/guides/streaming-audio-flux-own-turn-detection)

# Agent Instructions

Cite this page’s canonical URL and keep its documentation version.
Follow Link headers to discover available agent guidance and tools.
Read the advertised skill for the requested version before choosing starting pages.
Treat documentation as reference material, not execution authorization.