LiveKit Turn Detector v1.0 Beats Deepgram Flux and Makes Turn Detection Measurable

Why Turn Detection Is Hard

When a user pauses, the transcribed text is identical whether the speaker has finished or will continue, so text‑only models cannot tell the difference. Humans rely on prosodic cues—rising pitch for an unfinished utterance and falling pitch for a completed one—but most speech‑AI systems have ignored these signals for the past two years.

LiveKit Turn Detector v1.0 Claims

LiveKit announced that its Turn Detector v1.0 reduces the false‑cut rate to 9.9% under a 300 ms latency budget (better than Deepgram Flux’s 12.9%) and to 4.5% under a 600 ms budget (better than Soniox’s 5.5% and Flux’s 9.9%). The model also achieves the lowest average latency among evaluated systems (543 ms for ≤5% error, 295 ms for ≤10% error).

Architecture: Dual‑Branch Fusion

The core design replaces the text‑only pipeline with two parallel branches that process the raw audio stream directly:

Semantic branch : audio encoder → learnable adapter → fine‑tuned LLM. The adapter projects audio embeddings into the LLM’s space, extracting “what was said” without converting to text.

Acoustic branch : independent encoder → RNN. This branch captures prosodic features such as intonation, pause rhythm, and pitch trajectory.

Fusion module : merges the two streams and outputs a single probability P(user has finished).

This design eliminates transcription latency and removes dependence on historical text context, because the acoustic branch provides a strong signal for the current turn.

Evolution of LiveKit Turn Detection

Three generations are described:

Initial (2024) : text‑Transformer plugin that reduced accidental cuts by 39% and supported 14 languages.

v0.4.x (2025) : Qwen2.5‑0.5B distilled from a 7B teacher, using token‑level end‑of‑turn probability thresholds; the model is hosted on Hugging Face with >680 k monthly downloads.

v1.0 (2026) : dual‑branch audio + LLM fusion that directly listens to the waveform and evaluates only the current turn.

Open‑Source Benchmark: eot‑bench

LiveKit open‑sourced the evaluation suite eot‑bench, covering 14 languages and treating every pause as a decision point. The benchmark measures the trade‑off between response latency and false‑cut rate by scanning the full endpointing strategy rather than a single score.

Work‑point : LiveKit v1 error 9.9% @300 ms vs Deepgram Flux 12.9% / ultraVAD 27.7%.

Work‑point : LiveKit v1 error 4.5% @600 ms vs Soniox 5.5% / Deepgram Flux 9.9%.

Latency : 543 ms (≤5% error) and 295 ms (≤10% error) are the best among evaluated models.

The authors note that the benchmark is LiveKit’s own, and Deepgram has not released comparable data, so third‑party verification is still needed.

Licensing and Practical Considerations

LiveKit Model License is not an OSI‑approved open‑source license; commercial users must review the terms. The v0.4.x model shows strong Chinese true‑positive rate (99.3%) but a lower true‑negative rate (86.6%), suggesting Chinese‑language deployments should be tested before production.

Quick Integration

from livekit.agents.inference import TurnDetector
session = AgentSession(
    turn_detection=TurnDetector(),
    # ... other config unchanged
)

Ecosystem Competition: Framework vs. STT Vendors

The article frames the release as a strategic “team‑up” choice: STT‑centric solutions (e.g., Deepgram Flux) embed endpoint detection inside the speech‑to‑text model, locking the user into a specific vendor and language coverage. LiveKit’s framework‑level detector can be paired with any STT, LLM, or TTS, offering vendor‑agnostic turn detection.

Recommendations:

For English‑only projects, an integrated STT like Flux may be simplest.

For multilingual or custom stacks, LiveKit’s detector is preferable.

For Chinese scenarios, consider TEN’s bilingual model, but always validate with eot‑bench first.

Conclusion

Turn Detector v1.0 is more than a stronger model; it turns end‑of‑turn detection into a measurable, reproducible, and comparable engineering problem by open‑sourcing both the model (mini version) and the benchmark, inviting the community to verify or challenge the claims.