turn-1-mini

turn-1-mini

Detects when a speaker has finished talking and when someone is cutting in, using audio alone on a CPU in real time. It suits builders of voice agents and call tools.

Made by p99lab on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: turn-1-mini is a 5.4M-parameter model for conversation timing. It listens to audio and works out when a speaker has finished a turn and when another person is breaking in. It streams, uses audio only with no transcript, and keeps up in real time on a CPU. Version 2 adds support for several speakers.

Who it suits: It suits developers who build voice agents, meeting tools or call software that must decide when to speak and when to wait.

Industry | Who uses it | How they use it | Example
Voice AI | Voice agent developers | Decide when the agent may take its turn | Letting an agent reply once the caller stops
Meetings | Meeting tool builders | Track who holds the floor across several channels | Following four headset channels in a meeting
Call software | Contact center engineers | Spot interruptions and cut-offs in calls | Flagging when a speaker is being talked over
Research | Dialogue researchers | Compare turn-taking models on public benchmarks | Reading its scores against a published system

People who need a transcript or word-level timing can skip it, since it works from audio signals only.

Where to use it:

• Python with PyTorch on a CPU, on Windows, macOS or Linux
• ONNX streaming graphs, which the repository also holds for version 1
• Voice agents that receive each speaker's channel
• Call and meeting pipelines

Use case | Who | What they get back
End-of-turn detection | Voice agent developers | A decision that a speaker's turn is over
Interruption detection | Call software teams | A probability that a speaker is taking the floor
Speech activity per channel | Meeting tool builders | Voice activity and floor-claim signals for each channel

How to set it up:

• Python on your computer
• PyTorch, since version 2 runs with PyTorch on a CPU
• The p99turn package, which loads the model
• About 21 MB for the version 2 weights file

1. Install the Hugging Face hub package.
2. Download the model files.
3. Install the p99turn package with its torch extra, as the model card shows.

pip install -U huggingface_hub

hf download p99lab/turn-1-mini

How to use it: Feed the model audio from one or more speaker channels. With several channels, the model treats a turn as finished once that speaker is silent and another person starts claiming the floor. You can also pass the other channels, or your agent's own playback, as an optional REST input. Without it the model runs single-channel. The card's quick start holds the decision rules, which are not in the text provided here.

Tool | What it does
EOT output | Decides that a speaker's turn has ended
INT output | Gives the chance that this speaker is taking the floor, rather than saying yeah or laughing
REST input | Optional audio of the other channels or your agent's playback
Turn1Mini and Turn1MiniOnnx | Load the version 1 files as before

Example: Four people sit in a meeting, each on their own headset channel. You run all four channels through the model together. It marks when each person holds the floor, when their turn ends and when someone cuts in.

For newcomers and experts: Voice assistants often cut people off or wait too long. This model gives a quick signal for when a person is done speaking. A first small step is to download the files and read the quick start in the full card.

On the multi-party benchmark, the card reports 92.2 percent recall with 3.8 percent false alarms and a median delay of 210 ms. It also says it does not beat the VAP system on latency, and that its interruption output is not a strength on that benchmark.

Audio from each speaker's channel
Model scores speech and floor claims
Optional REST input from other channels
End-of-turn and interruption decisions

Good to know:

• License: Apache 2.0.
• Cost to run: none named. It runs on a CPU in real time, per the card.
• Data it reaches: it downloads files from the Hugging Face Hub, then processes the audio you give it.
• Not for: transcription. It works from audio only and produces no transcript.
• Version 1 files stay in the repository, unchanged.

Title: End-of-turn results on the multi-party test, v1 against v2
v1 recall: 86.9 percent
v2 recall: 92.2 percent
v1 false alarms: 25.3 percent
v2 false alarms: 3.8 percent

On Hugging Face, 13 November 2026

• Task: voice activity detection
• Downloads in the last 30 days: 2,733
• Likes: 6
• License: apache-2.0
• Languages: en
• Fine-tunes: 1
• Tags: onnx, turn-1-mini, turn-detection, end-of-turn, endpointing, turn-taking, interruption, barge-in, multi-party, voice-activity-detection, voice-agents, streaming

Links

• Model: https://huggingface.co/p99lab/turn-1-mini
• Files: https://huggingface.co/p99lab/turn-1-mini/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download p99lab/turn-1-mini

Then follow the usage in its model card. Nothing in this kit runs by itself.