# SPEV: Audio-Native Turn Detection

Purpose: Agent-readable summary of the SPEV landing page.
Last updated: 2026-10-19

SPEV is an open-weight (Apache-2.0) decision model from Sprag. It decides whether a speaker has finished in 166 ms median, straight from audio, with no transcript in between. Hosted SPEV calls are free on Sprag through October 31, 2026.

## Facts

- livekit/eot-bench accuracy: SPEV 77.56% vs. Whisper -> OpenAI 70.7%
- Median round-trip latency: 166 ms vs. 752 ms
- Est. price per 1,000 evaluations: $0.308 vs. $1.156
- Inputs: text, native audio, one inline image, or video frames. 8.2K-token context.

## Related links

- [Decisions docs](https://sprag.ai/docs/decisions.md)
- [SPEV model](https://sprag.ai/models/spev.md)
- [Weights on Hugging Face](https://huggingface.co/SpragAI/SPEV)