TurnSense

<div align="center"> <img src="./TenSense/image/Baiji_Team.png" alt="Baiji Team Logo" width="1000" height="450"/> <br/>

TurnSense

🎯 Lightweight · Accurate · Three-Class — Redefining Speech Turn Detection

<br/> <center><strong>47M 参数 ｜ CPU 延迟 ~55ms ｜ F1 高达 96.35% ｜ 无效语义过滤</strong></center> <br/>

</div> <br/>

Language: English | 中文

<br/>

⭐ If TurnSense is useful to you, please give us a Star! It helps us keep improving the model and documentation.

<br/>

📖 Table of Contents

• Why TurnSense
• Overview
• Key Features
• Model Size Comparison
• Benchmark Results
• Quick Start
• Evaluation Guide
• Citation
• Contact & Community
• License

<br/>

---

<br/>

🏆 Why TurnSense

<div align="center"> </div> <br/>

---

<br/>

📌 Overview

TurnSense is a three-class semantic detection model designed for human-machine voice interaction, focused on solving a critical problem in dialogue systems:

During a user's speech, should the system respond immediately, or continue waiting?

Traditional approaches typically rely on a simple binary classification — "finished or not." TurnSense goes further by simultaneously modeling semantic completeness and invalid input detection, enabling more natural turn-taking in complex real-world scenarios and significantly reducing false interruptions, premature responses, and noise-triggered activations.

<div align="center"> <img src="./TenSense/image/TurnSense.svg" alt="TurnSense Three-Class Illustration" width="820"/> </div> <br/>

TurnSense classifies user input into three semantic states:

These three labels enable the system to determine not only "should I respond?" but also "is it worth responding to?" — significantly improving interaction naturalness and system stability in voice assistants, real-time calls, intelligent customer service, and more.

<br/>

---

<br/>

✨ Key Features

🧠 Semantic-Level Three-Class Detection

Simultaneously models complete / incomplete / invalid states — closer to real conversational behavior than traditional binary classification, and currently the only open-source solution with native invalid utterance detection.

⚡ Ultra-Lightweight, Ultra-Fast Inference

Only 47M parameters (INT8 version ~50MB). CPU inference latency: p50 ≈ 54.65ms, p90 ≈ 58.00ms — meets the strict requirements of real-time interaction without a GPU.

🎯 Leading Accuracy

Achieves F1 96.35% (complete) and F1 96.32% (incomplete) on easyturn_real_test_ZH (300 samples), and F1 92.30% (complete) and F1 91.62% (incomplete) on semantic_test_ZH (2000 samples) — best or runner-up among all comparable models.

🚫 Invalid Input Filtering

On the NonverbalVocalization test set, invalid utterance precision reaches 100% with recall of 90.37% (F1 = 94.34%), effectively suppressing false triggers from non-verbal sounds and noise.

⚖️ More Robust Turn Decisions

Balances precision and recall in semantically ambiguous, pause-heavy, or colloquial scenarios, reducing both premature responses and missed responses.

📊 Reproducible Evaluation Framework

Ships with a complete evaluation pipeline and scripts, supporting unified metric comparison and performance regression analysis for full reproducibility.

🤗 Open-Source Friendly, Plug-and-Play

Standardized repository structure with FP32 / INT8 ONNX models — from installation to inference in just a few minutes.

<br/>

---

<br/>

📐 Model Size Comparison

<div align="center"> </div>

💡 With only 47M parameters, TurnSense achieves three-class capability — the best balance between accuracy and model size.

<br/>

---

<br/>

📊 Benchmark Results

All results below are based on open-source Chinese evaluation sets. Latency marked with (GPU) indicates GPU environment; otherwise, latency was measured on CPU.

<br/>

📋 easyturn_real_test_ZH (300 samples)

Data source: Real data samples from Easy-Turn-Testset

🔍 Key Finding: TurnSense achieves the highest F1 on both complete and incomplete classes, and is the only model with CPU p50 < 60ms while maintaining F1 > 96%.

<br/>

📋 semantic_test_ZH (2000 samples)

Data source: Chinese test split from KE-Team/SemanticVAD-Dataset

🔍 Key Finding: On the larger 2000-sample test set, TurnSense still maintains the best F1, demonstrating strong generalization capability.

<br/>

📋 NonverbalVocalization_invalid (728 samples)

Data source: OpenSLR Deeply Nonverbal Vocalization Dataset (SLR99)

🔍 Key Finding: TurnSense is currently the only model that supports invalid utterance detection. A precision of 100% means zero false positives — effectively preventing noise from triggering system responses.

<br/>

---

<br/>

🚀 Quick Start

1. Installation

git clone https://github.com/Baiji-Team/TurnSense.git
cd TurnSense

pip install -U numpy onnxruntime torch librosa soundfile pandas scikit-learn huggingface_hub

2. Model Weights

TurnSense model weights are available on Hugging Face: Baiji-Team/TurnSense

Download Options:

Option 1: Auto-download (Recommended) The inference script includes built-in Hugging Face download logic. The model will be automatically fetched and cached on first run.

Option 2: Git LFS

git lfs install
git clone https://huggingface.co/Baiji-Team/TurnSense

Option 3: Hugging Face Hub

from huggingface_hub import snapshot_download
snapshot_download(repo_id="Baiji-Team/TurnSense")

3. Inference

python infer.py

Example output:

Loading model from Baiji-Team/TurnSense...
Running inference on: "我想问一下那个订单就是昨天..."

Results:
  Input: "我想问一下那个订单就是昨天..."
  TurnSense Detection Result: "incomplete"

<br/>

---

<br/>

🧪 Evaluation Guide

1) Evaluation Pipeline

1. Load the .jsonl test dataset (line-by-line JSONL)
2. Warm up each model (default warmup_iters=20)
3. Run per-sample inference, collecting classification and performance metrics
4. Automatically generate summary and detail files

Output files include:

2) Data Format (JSONL)

Each line is a JSON object containing at least the following fields:

Example:

{"audio_path":"/001.wav","text":"帮我查一下明天上海天气","label":"complete"}
{"audio_path":"/002.wav","text":"我想问一下那个订单就是昨天...","label":"incomplete"}
{"audio_path":"/003.wav","text":"啊…嗯…（持续噪声）","label":"invalid"}

3) Run Evaluation

python TurnSense/Turn_benchmark/benchmark.py

<br/>

---

<br/>

📚 Citation

If you use TurnSense in your research or product, please cite:

@misc{turnsense2026,
  author       = {Baiji Team},
  title        = {TurnSense: A Three-Class Semantic Detection Model for Complete, Incomplete, and Invalid Utterances},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Baiji-Team/TurnSense}},
}

<br/>

❓ Contact & Community

If you have questions or suggestions, feel free to reach out:

<br/>

📄 License

This project is released under the Apache License 2.0 with certain additional conditions. See LICENSE for details.

<br/>

---

<div align="center">

Built with ❤️ by Baiji Team

</div>