From VAD to Flux: Built‑in End‑of‑Turn Eliminates Latency vs Interruption Trade‑off

When building real‑time voice agents, developers often encounter two opposite problems: the system either cuts the user off ("抢话") or pauses awkwardly for 800 ms–2 s ("尬停"). The root cause is turn detection – determining whether the user has finished speaking.

1. Why turn detection matters more than transcription accuracy

The speech‑agent pipeline (listen → transcribe → understand → decide → synthesize → play) is long. While word error rate (WER) is important, the perceived naturalness of a conversation depends heavily on the timing of hand‑overs. Humans pause, use filler words, and leave utterances unfinished; a system that only watches acoustic energy will misinterpret these pauses as the end of a turn, leading to three typical failures: cutting off the user, long dead air, and non‑reproducible behavior across noisy and quiet channels.

2. The old paradigm: VAD + silence threshold

Traditional real‑time ASR was designed for subtitles and meeting minutes, not for interactive agents. The common stack consists of:

ASR : produces the transcript.

VAD : decides whether speech is present.

Custom logic : if silence exceeds N ms, the turn is considered finished.

This approach is easy to explain, tune, and works with any cloud ASR for demos or low‑traffic scenarios. However, its ceiling is fixed: it uses acoustic energy as a proxy for semantic completion, so the threshold must be constantly re‑calibrated for different accents, speaking speeds, background noise, or when users think while speaking. The engineering debt grows as the product scales.

3. The new paradigm: ASR‑integrated End‑of‑Turn (EOT)

Deepgram’s Flux model treats turn detection as a first‑class task inside the ASR model. It learns both the transcription and the semantic cue that indicates whether the speaker has truly finished, when to respond, and how to keep the dialogue flowing. This removes the external detector layer and the need for hand‑crafted heuristics, while still allowing product teams to define interruption policies (soft, hard, barge‑in) on top of the model‑provided endpoint.

4. Independent validation by Coval

Coval’s simulation platform, which mimics production voice‑agent environments, measured Flux against Deepgram’s previous Nova‑3 model. The results show:

≈50 % lower first‑token latency.

Faster and more stable turn detection (median latency near 0 ms, competitors 0.5–1 s slower).

Comparable word‑error rate to Nova‑3.

Reduced tail latency, which is critical because occasional long pauses feel “stuck” to users.

These numbers demonstrate that the trade‑off between latency and interruption quality can be eliminated.

5. Choosing the right approach

The article proposes a three‑path selection framework:

Path A – VAD + silence : suitable for prototype, scripted, highly controlled or compliance‑driven scenarios where explainability is required.

Path B – Hybrid : keep the existing ASR + VAD stack but add a better turn‑model or rule‑based layer; useful when a vendor does not yet provide built‑in EOT but the product cannot be fully rewritten.

Path C – Built‑in EOT (Flux‑class) : ideal for open‑domain, high‑interaction use cases (sales assistants, medical appointments, verification dialogs) where natural turn‑taking is the north‑star metric.

Selection criteria include configurable silence thresholds, scene‑specific tuning (quiet headset vs open office), and tracking metrics such as first‑token P50/P95 latency, mis‑interrupt rate, missed‑turn rate, and cost.

6. Product‑level considerations

Even with a high‑quality EOT, the product layer must still decide how to handle user barge‑in, hard interruptions in emergencies, and more conservative endpoints for confirmation dialogs. The recommendation is to split responsibilities: the model supplies accurate endpoints and transcripts; the strategy layer enforces business‑level interruption policies. Mixing the two can make a good model appear reckless.

7. Looking ahead

Flux is positioned within a broader narrative of sharing dialogue context across STT, LLM, and TTS components, moving beyond flattening speech into plain text. Short‑term actions: replace silence‑timer‑based turn detection with model‑based EOT. Mid‑term: enable the entire speech stack to share dialogue state so that tone, intent, and turn information flow seamlessly.

8. Takeaway

In 2026 the default assumption for conversational speech shifts from “detect voice energy” to “understand when a turn truly ends.” Teams should evaluate their own recordings, measure latency distribution, mis‑interrupt and missed‑turn rates, and ask three questions: Do our users pause like real conversation? Is our north‑star metric transcription accuracy or natural hand‑over? Have we validated turn quality in our own scenario?