Migrating from Nova-3 to 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

• Encoding: See Audio Format Requirements table below
• Sample rates: See Audio Format Requirements table below
• Channels: Mono only
• Chunk size: 80ms strongly recommended for optimal model performance and latency.

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

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

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

{
  "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.

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

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

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.

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.

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

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.

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

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 message when your own detector fires. See Bring Your Own Turn Detection for the full recipe.

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

• 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; see Bring Your Own Turn Detection