Open weights · Apache-2.0 · Free through Oct 31
SPEV: OmnimodalJev for audio
Trained for turn detection, SPEV knows when a speaker is done in 166 ms, straight from audio. Ask it any other question about audio the same way and get an answer with a probability for every option, no transcript required.
- 166 ms median decision
- 77.56% on livekit/eot-bench
- Open weights, Apache-2.0
Transcribe-then-decide is slower and loses the signal.
The usual way
~752 ms
With SPEV
166 ms
Accuracy
SPEV 77.56% vs. Whisper → OpenAI 70.7% (higher is better).
Median round-trip latency
SPEV 166 ms vs. Whisper → OpenAI 752 ms (lower is better).
Est. price per 1,000 evaluations
SPEV $0.308 vs. Whisper → OpenAI $1.156 (lower is better).
Your first decision in under five minutes
- 1Create a free account
- 2Copy your API key
- 3Send audio with a question
{
"model": "spev",
"input": "",
"input_audio": { "format": "wav", "data": "BASE64_ENCODED_AUDIO" },
"questions": [
{
"type": "choice",
"name": "turn",
"instructions": "Has the most recent speaker yielded the floor or are they continuing their turn?",
"choices": [
{ "value": "keep_listening", "description": "The speaker is continuing their turn." },
{ "value": "respond", "description": "The speaker has yielded the floor; another speaker can respond." }
]
}
]
}Free through October 31. Usage-based after.
Putting SPEV in front of real users? Talk to us.
Hosted at scale
Volume pricing and rate limits for production traffic.
Self-host the open weights
30B parameters, built on Qwen3-Omni-30B-A3B with a trained LoRA adapter and decision head, Apache-2.0. Self-hosting takes some setup today, and we can help.
Weights on Hugging Face →Tune the decision
Design questions and thresholds for your product's turn-taking.
FAQ
Free for 13 more days. Make your first call today.
Sprag
Lightning-fast inference infrastructure for production-grade multimodal models.