TEN Turn Detection

Intelligent model for detecting conversation turns in human-AI dialogue systems, identifying three key utterance states for natural conversation flow.

Core Functionality

Classifies user text into three states:

• Finished: Complete thought expressed; user expects a response
• Unfinished: Clearly unfinished utterance where user pauses but intends to continue
• Wait: User requests AI to pause or stop speaking

Technology

• Model: Transformer-based (Qwen2.5-7B)
• Languages: English and Chinese
• Inference: Production-ready deployment scripts included

Performance Metrics

Strong accuracy across all detection categories:

Performance exceeds comparable open-source alternatives.

Use Cases

• Full-Duplex Conversation — Enable interruption detection and natural turn-taking
• Conversation Flow Management — Determine when to wait for user continuation vs. respond
• Pause/Stop Handling — Recognize when users request the AI to pause or stop
• Multi-Language Support — Deploy in English and Chinese speaking applications

Getting Started

# Basic usage
from ten_turn_detection import TurnDetection

detector = TurnDetection()
state = detector.detect("user input text")
# Returns: "finished", "unfinished", or "wait"

Integration with TEN Agent

TEN Turn Detection works seamlessly with TEN VAD for complete conversation management:

1. TEN VAD detects speech activity (when user is speaking)
2. TEN Turn Detection analyzes utterance to determine conversation state
3. Agent responds appropriately based on detected turn state

This enables natural, responsive voice conversations with proper turn-taking.

Datasets

Bilingual test datasets (TEN-Turn-TestSet) are available for:

• Validation and testing
• Fine-tuning on domain-specific language
• Benchmarking model performance

Resources

• GitHub Repository: TEN-framework/ten-turn-detection
• Model Weights: Available via Hugging Face
• Datasets: Bilingual test datasets included

License

Apache 2.0 (see repository for details)