🌐 CoughChain: An Environmentally Adaptive Cough Detection Framework Guided by Diverse Chain-of-Thought Prompting in LLM

CoughChain is a training-free, large language model (LLM)-based cough detection framework that leverages diverse chain-of-thought (CoT) prompting strategies and environment-aware context injection to achieve robust and interpretable cough recognition in complex acoustic environments.

🔍 Core Idea: By explicitly informing the LLM whether the input audio comes from a "quiet" or "noisy" environment—and guiding its reasoning through structured CoT prompts—the model adaptively adjusts its decision logic without any fine-tuning or parameter updates.

---

📌 Key Highlights

• ✅ Zero training required: Uses off-the-shelf LLM (Qwen-Omni-Turbo) via prompt engineering only
• 🧠 Four CoT reasoning strategies:
  • audiofeature: Audio feature–guided reasoning (e.g., spectral, energy, zero-crossing rate)
  • step: Linear, step-by-step logical decomposition
  • selfask: Iterative self-questioning and answering to refine judgment
  • tot (Tree of Thought): Multi-branch tree-like reasoning with path exploration and consensus
• 🌍 Environment-adaptive: Dynamically switches reasoning strategy based on acoustic context ("quiet" vs. "noisy")
• 🎯 Evaluated on a balanced, real-world multi-scenario cough dataset

---

📂 Code Structure

The repository contains three main directories, each corresponding to a specific experimental setting:

Directory / File	Description
cot/	Baseline & four CoT implementations using Qwen-Omni-Turbo: • audiofeature.py – LLM reasoning guided by audio features • step.py – Stepwise logical decomposition of cough judgment • selfask.py – Self-questioning to iteratively focus on key evidence • tot.py – Tree-based multi-path reasoning with final consensus
quiet/	Quiet-environment evaluation: Runs CoughChain with "quiet" context injected into prompts
noisy/	Noisy-environment evaluation: Runs CoughChain with "noisy" context to enhance robustness against interference

💡 All implementations are purely prompt-driven—switching environments or reasoning styles only requires modifying the prompt template. No model weights are altered.

---

⚙️ Essential Setup Before Running

Before executing any script, complete the following two critical configurations:

1. Configure Your Qwen API Key

Obtain your Qwen-Omni-Turbo API key from the DashScope Console, then choose one of the following methods:

• Option A: Hardcode in script (for quick testing)

  client = OpenAI(
      api_key="YOUR_API_KEY_HERE",  # ← Replace with your actual key
      base_url="https://dashscope.aliyuncs.com/..."
  )

2. Update the Audio Root Directory Path (❗Critical!)

In every Python script (e.g., cot/audiofeature.py, noisy/selfask.py, etc.), locate the following line:

root_audio_dir = "/path/to/your/cough_audio_dataset"  # ← MUST UPDATE!Replace it with the absolute or relative path to your local cough audio folder.

📊 Code and Data Availability

• 🔒 The full source code will be publicly released upon official acceptance of the paper.
• 📁 A subset of the cough audio dataset will be provided in this repository .
• 📩 For access to the complete dataset (including noisy-scenario recordings and metadata), please contact us via email at: fmingkk@163.com.

We encourage the use of this data strictly for academic and non-commercial research purposes, in accordance with ethical guidelines.