🦠 SVM-Based Cough Detector

Welcome to the SVM-Based Cough Detector project! This tool uses machine learning to automatically detect coughs in audio recordings.

🚀 What It Does

1. Loads and Segments Audio: Takes in .wav files and fetch positive (cough) segments based on the labels, and negative (not cough) segments that has the similar length distribution as positive samples.
2. Extracts Features: Pulls out MGCC features from each segment and aggregates them with mean and standard deviation to create uniform feature vectors.
3. Trains an SVM Model: Uses extracted features to train a Support Vector Machine (SVM) that can differentiate between cough and non-cough sounds.
4. Predicts coughs in audio collected by myself: Applies the trained model to new audio files to identify and timestamp cough events.
5. Store results: Generates label files where coughs are detected in the audio waveform.

📋 How to inference

1. Clone the Repository:

   git clone https://github.com/yourusername/svm-cough-detector.git
   cd svm-cough-detector

2. Install Dependencies: (recommend to do it in a virtual enviroment)

   pip install -r requirements.txt

3. Inference:

   python3 src/inference.py --model svm --input-dir dir/contains/one/data.wav

4. Want to record your own audio files?

   python3 src/record_audio.py --output-dir dir/to/save/the/recording --duration 5

• File type: The audio file is saved in .wav formate with sample rate at 16KHz and is monophonic sound.
• File name: The audio file is saved as data.wav

5. Usage example:

   # Record a 5 second audio and save it to my_data/positive/cough_example
   python3 src/record_audio.py --output-dir my_data/positive/cough_example --duration 5
   
   # Use the recorded audio for inferencing
   python3 src/inference.py --model svm --input-dir my_data/positive/cough_example
   

🛠️ How It Works

1. Audio Preprocessing

• Loading: Utilizes librosa to load audio files at a consistent sampling rate.
• Segmentation: Splits audio into positive and negative segments. During this step, we try to keep both side has simialr distrubution in length and number of samaples.

2. Feature Extraction

• MFCCs: Extracts Mel-Frequency Cepstral Coefficients (MFCCs) from each segment.
• Aggregation: Calculates the mean and standard deviation of MFCCs to create a fixed-length feature vector for each segment.

3. Model Training

• Scaling: Standardizes features using StandardScaler for better SVM performance.
• SVM Training: Trains an SVM classifier with hyperparameter tuning using GridSearchCV to find the best settings.

4. Prediction Pipeline

• Segmenting New Audio: Processes new .wav files under my_data folder. Breaks audio signal down into fixed-size (0.1s) chunks
• Feature Processing: Extracts and scales features from new segments.
• Classification: Predicts whether each segment contains a cough.
• Mapping: Associates predictions with their corresponding time frames and filters out non-cough segments.