What turn-1 decides
Voice agents either cut people off or leave them waiting. At every pause, turn-1 decides whether the speaker has finished or is only thinking, so your agent answers when it should.
# example call
00:02.1 speech "I'd like to move my appointment to…"
00:03.4 pause hold not finished
00:04.2 speech "…Thursday afternoon, if you have it."
00:05.6 pause end_of_turn agent repliesTwo models
turn-1-mini
Small and streaming: end of turn and interruptions.
- Parameters
- 5.3M
- Size
- 21 MB · 5.4 MB with 8-bit weights
- ONNX build
- 7 MB, runs without PyTorch
- Licence
- Apache-2.0
turn-1
The larger end-of-turn model, for servers.
- Runs on
- Servers
- Plug-ins
- Pipecat · LiveKit
- Licence
- Apache-2.0
How they are built
turn-1-mini is entirely our own: a 5.3M-parameter streaming model that we designed and trained. It contains no third-party weights.
turn-1 is a turn-detection model by p99lab, built on open speech encoders. The turn-detection model, its data pipeline and its training are ours.
Where it runs
turn-1 runs in your own stack, next to any speech recognizer, language model and voice. Use the plug-ins for Pipecat and LiveKit, or load the weights or the ONNX build directly.
With Audrey
Audrey's English streams run turn-1-mini on the server and send turn_end when the speaker has
finished. See turn detection.
Results
Evaluation results and how they were measured are on each model card on Hugging Face.